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

Abstract

The application 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: preprocessing and text conversion of original node attributes of a dynamic network structure graph; using a large language model (LLM) to analyze the converted dynamic network structure graph; performing feature extraction from a semantic embedding side and a structure embedding side respectively; performing semantic embedding and structure embedding feature fusion by using a multi-head attention mechanism; and performing a downstream task based on the fused embedding to realize data monitoring. The LLM-GNN collaborative optimization system proposed by the application realizes intelligent breakthrough of marine oil and gas system data monitoring.

Description

Technical Field

[0001] This invention relates to the field of marine environmental monitoring technology, and in particular to a method and system for monitoring marine oil and gas equipment data based on augmented graph learning. Background Technology

[0002] As deep-sea oil and gas development extends to ultra-deep water areas above 1500 meters, the system faces three major technical bottlenecks: First, the challenge of processing ocean-specific data. Digital nodes in deep-sea oil and gas fields form a three-dimensional monitoring network covering thousands of square kilometers, with unstructured data accounting for over 65%. Traditional numerical analysis struggles to resolve the nonlinear noise interference unique to the marine environment. Second, insufficient accuracy in dynamic modeling. Existing static models cannot adapt to accelerated aging under high pressure and low temperature environments, the synergistic effects of seawater corrosion and microbial adhesion, and dynamic load changes during typhoon seasons, leading to a 18-25% increase in false alarm rates in scenarios such as deep-water pipeline stress monitoring. Third, the requirement for real-time response in extreme environments. For high-risk scenarios such as blowout warnings and submarine landslide monitoring, a decision-making loop from data acquisition to emergency shutdown must be completed within 300 milliseconds. However, current multimodal data processing latency cannot meet the millisecond-level response requirements in areas with strong ocean currents.

[0003] Currently, while graph neural network (GNN)-based data analysis methods have made some progress in equipment relationship modeling and data monitoring, their application in real-world scenarios still faces several key technical bottlenecks. Specifically, existing GNN solutions mainly face the following challenges: First, in terms of topology modeling, traditional methods generally adopt the assumption of a static graph structure, which cannot effectively capture the essential characteristics of the dynamic evolution of equipment nodes. Second, in terms of data utilization efficiency, existing systems suffer from severe modal imbalance. Industry survey reports indicate that approximately 78% of industrial monitoring systems still remain at the level of traditional numerical signal processing, simply archiving or discarding textual data such as equipment logs and maintenance records. This processing method directly leads to the loss of key semantic information such as equipment operating context and maintenance history, making it impossible for the model to identify complex fault modes. Furthermore, in terms of multimodal fusion methods, current mainstream shallow fusion strategies such as feature concatenation or ensemble voting are insufficient to establish a deep correlation between semantic features and physical signals. This superficial fusion approach fails to capture cross-modal causal relationships such as "insulation aging and partial discharge signal enhancement described in logs," ultimately leading to significant deviations between generated maintenance recommendations and actual conditions. These limitations severely restrict the actual effectiveness of marine oil and gas equipment systems, necessitating breakthroughs through technological innovation in the following areas: real-time evolution capabilities of dynamic topologies, deep semantic understanding of multimodal data, collaborative reasoning mechanisms for cross-modal features, and improvements in industrial-grade computing efficiency. Only by addressing these issues can the application potential of GNNs in the field of intelligent equipment operation and maintenance be truly unleashed.

[0004] Meanwhile, existing methods combining Large Language Models (LLM) and Graph Neural Networks (GNN) face the following challenges in practical applications: First, LLM excels at handling semantic information but lacks the ability to model graph structures, while GNN, although capable of capturing topological relationships, struggles to understand complex semantics, leading to difficulties in feature alignment during fusion. Second, data in industrial network scenarios (such as equipment status and user behavior) often change dynamically, making it difficult for existing methods to adaptively adjust graph structures or efficiently fuse multimodal inputs, thus limiting the model's generalization ability. Furthermore, the large number of parameters in LLM, coupled directly with GNN, leads 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. Additionally, both LLM and GNN suffer from black-box issues, with their joint decision-making processes (such as node classification and anomaly detection) lacking interpretability, making it difficult to meet the reliability requirements of industrial scenarios. Finally, existing methods often employ fixed architectures, making it difficult to dynamically adjust the interaction between LLM and GNN for different tasks (such as fault diagnosis and supply chain optimization), leading to performance degradation during cross-task migration. Therefore, there is an urgent need for an efficient, interpretable, and dynamic LLM-GNN fusion method to solve the above problems. Summary of the Invention

[0005] To address the aforementioned problems, this invention provides a method and system for monitoring marine oil and gas equipment data based on augmented graph learning.

[0006] In a first aspect, the present invention provides a data monitoring method for marine oil and gas equipment based on enhanced graph learning, which adopts the following technical solution: A data monitoring method for marine oil and gas equipment based on reinforcement graph learning, comprising: Acquire multi-source sensor data and construct a dynamic network structure diagram with spatiotemporal characteristics; Preprocessing and text-based conversion of the original node attributes of the dynamic network structure graph; The transformed dynamic network structure diagram is analyzed using the Large Language Model (LLM). Feature extraction is performed from both the semantic embedding side and the structural embedding side. Utilizing a multi-head attention mechanism to fuse semantic embedding and structural embedding features; Data monitoring is achieved by embedding downstream tasks into the system.

[0007] Furthermore, the acquisition of multi-source sensor data and the construction of a dynamic network structure graph with spatiotemporal characteristics include: real-time acquisition of device operating status data by deploying a multi-source heterogeneous sensor network; construction of a dynamic network structure graph G=(V,A,X,Y) with spatiotemporal characteristics; expansion of the basic node set V; upgrading each device node vi∈V to an enhanced node entity with spatiotemporal awareness; upgrading the adjacency matrix A from a static definition to a spatiotemporal correlation matrix; and integrating the feature matrix X with four types of modal features to form a unified representation.

[0008] Furthermore, the preprocessing and text-based transformation of the original node attributes of the dynamic network structure graph includes mapping continuous attributes to the [0,1] interval using a min-max normalization method; converting discrete attributes into binary vectors using one-hot encoding; and generating a text fragment for each node, which contains all attribute information of that node, and sorting and describing the attributes according to the weights of semantic association, as follows: The attributes of node [node number] include: [attribute1 name]: [attribute1 value], [attribute2 name]: [attribute2 value], ..., [attribute n name]: [attribute n value]; Among them, the attribute names and attribute values ​​are sorted according to the weight of semantic association, with the attribute with the higher weight appearing first.

[0009] Furthermore, the method of using a Large Language Model (LLM) to analyze the transformed dynamic network structure graph includes analyzing the node attribute text using LLM to infer the potential relationships between nodes. For each pair of nodes, its attribute text is input into the LLM, and the relationship description output by the LLM is obtained. The relationship weight between each pair of nodes is calculated based on the similarity between the node and node attribute text vectors, as expressed below: , in, This represents the weight of the relationship between node i and node j. and Let i and j 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 method of using a large language model (LLM) to analyze the transformed dynamic network structure graph also includes extracting the topological features of underwater equipment nodes using a graph computation method based on adjacency matrices. Specifically, for the deep-sea production network, the degree centrality of each node is calculated as a basic structural attribute, and a dynamically weighted PageRank algorithm is used to accurately evaluate the global importance of underwater equipment nodes. This is achieved by introducing time-varying communication weights and environmental disturbance factors to adapt to the complex working conditions of the deep sea. The iterative calculation process is expressed as follows: , Where N ( i） This indicates the nodes constructed by the marine oil and gas equipment system. you The neighbor set, damping coefficient α ∈(0,1) is optimized to 0.8 based on the characteristics of the marine environment. N The total number of all built device nodes. These are time-varying edge weights, reflecting the node's... you and vj At any moment t Real-time communication load, It is an environmental disturbance factor. Set it to 0.8.

[0011] Furthermore, the method of using a Large Language Model (LLM) to parse the transformed dynamic network structure graph also includes using LLM for edge adjustment and adaptive graph structure optimization. First, LLM is used to parse node pairs (…). you , vj The text description of the relation is output, along with the relation type and confidence level, and a time-series relation tensor is constructed. Historical information is aggregated using LSTM, and then topological and semantic features are dynamically fused, represented as follows: , Among them, the balance factor β Learning jointly from topological and semantic features, .

[0012] Furthermore, feature extraction is performed from both the semantic embedding side and the structural embedding side. This includes, on the semantic embedding side, performing hierarchical semantic encoding on the construction nodes of the marine oil and gas system; for equipment operating at a specific time, extracting features within local time windows; dividing equipment logs into blocks according to fixed time windows, inputting them into a lightweight LLM to obtain local semantic representations, and employing a sliding window strategy to cover all time-series data; for continuously operating equipment, performing global time-series feature extraction; inputting the entire equipment historical log into an LLM to extract global context representations, designing dynamic gating, and fusing local and global features, represented as follows: , in, For the Sigmoid function, These are learnable parameters.

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

[0014] Furthermore, the multi-head attention mechanism for semantic and structural embedding feature fusion includes identifying and distinguishing embedding sources based on adversarial cross-modal alignment, performing linear transformations on semantic and structural embeddings respectively to generate query vectors, key vectors, and value vectors, calculating attention weights between semantic and structural embeddings to measure their importance in the fusion process, and using residual connections to fuse semantic information back into the original features to avoid information loss due to direct overwriting. Finally, downstream tasks are executed based on the fused embeddings, and the 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] Secondly, a data monitoring system for marine oil and gas equipment based on augmented graph learning includes: The data acquisition module is configured to acquire multi-source sensor data and construct a dynamic network structure diagram with spatiotemporal characteristics. The preprocessing module is configured to preprocess and convert the raw node attributes of the dynamic network structure graph into text. The parsing module is configured to parse the transformed dynamic network structure graph using the Large Language Model (LLM). The feature extraction module is configured to extract features from both the semantic embedding side and the structural embedding side. The feature fusion module is configured to perform semantic embedding and structural embedding feature fusion using a multi-head attention mechanism; The monitoring module is configured to perform downstream tasks based on fusion embedding to achieve data monitoring.

[0016] Thirdly, the present invention provides a computer-readable storage medium storing a plurality of instructions adapted for loading and execution by a processor of a terminal device of the aforementioned method for monitoring marine oil and gas equipment data based on augmented graph learning.

[0017] Fourthly, the present invention provides a terminal device, including 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, the instructions being adapted to be loaded and executed by the processor to provide the aforementioned method for monitoring marine oil and gas equipment data based on augmented graph learning.

[0018] In summary, the present invention has the following beneficial technical effects: (1) This invention innovatively integrates Large Language Model (LLM) and Graph Neural Network (GNN). Through the powerful semantic understanding and reasoning capabilities of LLM, it breaks through the limitation of traditional marine oil and gas equipment system data monitoring relying solely on structured data. LLM can deeply parse 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 transform them into dynamic weights that are injected into the adjacency matrix of GNN. This enables the graph structure, which could only model explicit physical connections, to capture implicit fault propagation paths. In actual tests in multiple industrial scenarios, this integrated architecture can effectively improve the accuracy of fault diagnosis.

[0019] (2) The LLM-GNN collaborative optimization system proposed in this invention achieves a breakthrough in intelligent monitoring of marine oil and gas system data. The system uses LLM to parse equipment status description text in real time, and combines Granger causality test and dynamic time warping algorithm to automatically identify potential equipment correlations and dynamically optimize the graph structure: when a significant causal relationship is detected between a motor current waveform and a bearing temperature sequence, 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 GNN for multi-level feature aggregation, and then the semantic embedding extracted by LLM and the structural embedding generated by GNN are dynamically fused through cross-modal attention mechanism.

[0020] (3) The system design of this invention has high adaptability and scalability to different scenarios. 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 graph by parsing order logs and logistics records, while GNN predicts potential delivery risk nodes; in the energy management scenario, LLM analyzes equipment energy efficiency reports to identify abnormal energy consumption patterns, while GNN establishes a network of connections between energy-consuming devices to locate potential energy-saving points. This flexible architecture allows enterprises to complete the model deployment of new scenarios in a short time by only adjusting the prompt word template and task-specific header (TaskHead), saving implementation costs compared to traditional customized development solutions. Attached Figure Description

[0021] Figure 1 This is a schematic diagram of a marine oil and gas equipment data monitoring method based on augmented graph learning according to Embodiment 1 of the present invention. Detailed Implementation

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

[0023] Example 1 Reference Figure 1 This embodiment of a method for monitoring marine oil and gas equipment data based on augmented graph learning includes: S1. Collect multi-source sensor data (numerical, text, image, etc.) from marine oil and gas equipment systems, construct a dynamic graph structure, in which the equipment is the node, the physical connection, communication relationship and semantic association between the equipment are the edge, and the node features are fused with numerical signals, text logs and spatiotemporal information to form a multimodal graph representation.

[0024] S2. Preprocess and text-convert node attributes, normalize and encode numerical features (such as sensor readings), perform One-Hot encoding on discrete features, and extract key semantic information from equipment logs to generate structured-text hybrid node feature representations, thereby enhancing the model's ability to understand complex industrial data.

[0025] S3. The device logs and status descriptions are parsed using a Large Language Model (LLM). The semantic information between nodes is output by LLM through a Prompt template, and the dynamic edge weights are calculated in combination with the original graph structure. At the same time, a temporal sliding window mechanism is introduced, and the adjacency matrix is ​​dynamically updated using LSTM.

[0026] S4. On the semantic embedding side, a hierarchical encoding strategy is adopted to extract local time window features and global temporal features from 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. Furthermore, adversarial training is introduced to force semantic embedding and structural embedding to align in the latent space through a domain discriminator, thereby enhancing the synergy of bimodal representations and providing more robust feature representations for downstream tasks.

[0027] S5. Design a multi-head attention mechanism to perform 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 joint embedding representation, and enhance the model's ability to model complex industrial scenarios.

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

[0029] Step S1 specifically includes: S1.1: In the marine oil and gas data twin intelligent monitoring system, a dynamic network structure graph G=(V,A,X,Y) with spatiotemporal characteristics is constructed 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.

[0030] The node set V = {v_1, v_2, ..., v_N} fully represents the various 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 spatial coordinates and installation location. The adjacency matrix A ∈ R^(N×N) accurately describes the complex relationships between nodes through three dimensions: hard relationships based on physical connections (such as cable connections and communication links), soft relationships based on data correlations (such as covariance analysis), and semantic relationships 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 the 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 by BERT), and 3) inherent device attributes (such as static parameters such as model and service life). The device tag set Y can be 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), and supports multi-task joint learning.

[0031] S1.2: To address the core issues of traditional static graph models' inability to handle spatiotemporal heterogeneity and multimodal fragmentation in marine oil and gas equipment system scenarios, the basic node set V is first intelligently extended, upgrading each device node vi∈V into an enhanced node entity with spatiotemporal awareness capabilities. Its data structure is defined as follows: vi={ ID: Unique identifier for the device (device code) Equipment type: A three-level classification system based on function (sensors / controllers / actuators). Physical coordinates: (x, y, z) three-dimensional spatial position (accurate to the millimeter level). Timestamp queue: {t1,...,tn} (records key state change times) Dynamic attributes: Sampling frequency: Device-specific sampling rate (e.g., 10Hz vibration sensor), Communication protocols: Modbus / TCP and other industrial protocol identifiers. Health status: Status coding based on ISO standards } To address the common problem of non-uniform sampling data in industrial settings, interpolation is used to unify the time granularity. , Assign hierarchical labels based on device type (sensor / controller / actuator) for subsequent attention in the hierarchical graph. (1) Sensor node: labeled as L1 level (data acquisition layer) (2) Controller node: labeled as L2 level (logic control layer) (3) Actuator node: labeled as L3 level (physical execution layer). Different attention calculation strategies are set between each level, 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 The static definition has been upgraded to a spatiotemporal correlation matrix, where i and j are node indices: , in These are learnable parameters, and the contribution of each mode is dynamically adjusted through backpropagation.

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

[0034] (2) Semantic association weight : By parsing device logs using LLM to generate relationship descriptions and calculating 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 for equipment status change).

[0035] (3) Causal time series weights : Alignment of Granger causality test with DTW time series: , The value is set to 1 if the Granger test p-value is less than 0.05 (indicating causality), and 0 otherwise. This metric ensures that only statistically significant causal relationships are preserved. In real-world industrial scenarios, lag order constraints are added to avoid long-delay spurious causality. The DTW time series alignment part addresses the issue of inconsistent sampling rates between devices and calculates the optimal alignment path cost between the two time series.

[0036] S1.4: Perform multimodal fusion of node features X: Feature matrix X Integrating the features of the four modalities to form a unified representation: , Temporal alignment: Sliding window and cubic spline interpolation are used to unify timestamps for non-uniformly sampled data; Spatial alignment: Device coordinates are mapped to a unified spatial grid (accuracy ±5cm) through Geohash-3D encoding; Semantic alignment: Text features are extracted using an industrially fine-tuned BERT model to construct an embedding space for device fault codes.

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

[0038] S2.2: Generate a text fragment for each node, containing all attribute information for that node, and sort and describe the attributes according to the weight of semantic association. Use the following template to generate the text description: The attributes of node [node number] include: [attribute1 name]: [attribute1 value], [attribute2 name]: [attribute2 value], ..., [attribute n name]: [attribute n value]. Among them, the attribute names and attribute values ​​are sorted according to the weight of semantic association, with the attribute with the higher weight appearing first.

[0039] 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, thereby optimizing the graph structure.

[0040] First, node relationships need to be inferred. LLM is used to analyze the node attribute text to infer potential relationships between nodes. For each pair of nodes, its attribute text is input into the LLM, and the relationship description output by the LLM is obtained. The following template is used to generate the relationship description: The relationship between node [node number 1] and node [node number 2] is: [relationship description]. S3.2 Calculate the relationship weight between each pair of nodes using the following formula: , in, This represents the weight of the relationship between node i and node j. and Let i and j 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, using a graph computation method based on adjacency matrices to extract the topological features of subsea equipment nodes. For key equipment nodes in the subsea production control system (such as subsea control modules (SCM), subsea valves, and jumpers), their degree centrality is first calculated as a fundamental structural property: for each node... Calculate its degree centrality as a fundamental structural property: , In deep-sea production networks, this indicator directly reflects the density of physical / data connections to underwater equipment. Typical manifestations include: High-value nodes (degree centrality > 15) are usually underwater master control stations or seabed distribution units (SDUs). Medium-value nodes (3 < degree centrality ≤ 15) are mostly relay equipment such as subsea production trees and manifolds; Low-value nodes (degree centrality ≤ 3) correspond to edge sensors (such as seabed pressure thermometers and corrosion monitoring probes). When the adjacency matrix contains edge weights When the weighted degree centrality is based on acoustic communication frequency or umbilical cable data transmission volume, the result is: , This improved indicator 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 in cross-connection rooms) 3) emergency communication paths (such as sudden communication loads at ROV docking interfaces).

[0042] Meanwhile, a Dynamically Weighted PageRank (DWPR) algorithm is employed to more accurately assess the global importance of underwater device nodes. This algorithm, based on the traditional PageRank algorithm, introduces time-varying communication weights and environmental disturbance factors to better adapt to complex deep-sea conditions. Its iterative calculation process is as follows: , Where N ( i） This indicates the nodes constructed by the marine oil and gas equipment system. you The neighbor set, damping coefficient α ∈(0,1) is optimized to 0.8 based on the characteristics of the marine environment (default value is 0.85). N This represents the total number of all built device nodes.

[0043] These are time-varying edge weights, reflecting the node's... you and vj At any moment t Real-time communication loads (such as acoustic data transmission volume, hydraulic signal strength); These are environmental disturbance factors used to quantify the impact of ocean currents on underwater wireless communication (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 it to 0.8 to match the latency characteristics of underwater communication.

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

[0045] First, the continuous signals (such as temperature and vibration amplitude) collected by the sensors are subjected to adaptive normalization based on operating conditions: , The sliding window W is dynamically adjusted based on the device type: Rapid response devices (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 employed to compensate for potential influences (such as acoustic communication delay, the non-periodicity of ROV inspection data, and sensor drift caused by tidal cycles), generating 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 handle asynchronous data alignment issues), and N is the number of particles that affect computational efficiency and the nonlinear characteristics of deep-sea data.

[0046] A sliding window (e.g., a 10-second window) is used to align sensor data at different sampling rates. After filling in missing values ​​using Kalman filtering, the data is concatenated along the feature dimension to form a composite matrix, followed by a concatenation operation. , For discrete features, perform enhancement operations: , Among them, LLM embedded input dimension dLLM =768 (BERT-base model), attention weighting is used for some special data depending on the situation. α ∈[0.8,1.5]). The standard clauses are encoded into vectors, and the topological structure of the standard document is extracted using a graph encoder.

[0047] Then cross-modal gating fusion is performed: , Among them, topological features Number of convolutional layers in the graph L =3: Covers the three-level topology of the subsea production system: "wellhead-manifold-umbilical cable"; Gating weight g: bias term b g Initialized to -0.5: Prefers prioritizing trust in topological features (due to the sparsity of marine text data). Threshold adjustment after activation: Forced under typhoon mode. g >0.7 (primarily based on topological stability).

[0048] S3.3 Edge adjustment using LLM: First, use LLM to resolve node pairs ( you , vj The text description (from S2.3) of the relation, outputting the relation type (causal / cooperative / dependent) and confidence level: , Constructing temporal relation tensors Historical information is aggregated using LSTM: , Next, the topological features and semantic features are dynamically fused: , Among them, the balance factor β Learning jointly from topological and semantic features: , S3.4 Adaptive Graph Structure Optimization: Dynamically adjust the original adjacency matrix : , Learning rate λ Adaptive adjustment of variance via sliding window: , The updated adjacency matrix needs to be sparsified (by pruning and edge addition) to increase its information density, including removing edges. ( (The value is set to 0.2), and for high-confidence semantic relations And topologically sparse nodes Add a new edge. Then, the oscillating edge needs to be adjusted. Apply smoothing constraints: , Step S4 specifically includes: S4.1 Semantic Embedding Learning (LLM Side) After adjusting the graph structure, semantic embedding and structural embedding are needed to facilitate subsequent optimization. These two types of embedding represent nodes from different perspectives, solving two key problems in graph data analysis: (1) Semantic embedding: The node attribute text is encoded using LLM, and semantic embedding captures the semantic information of node attributes. (2) Structural embedding: The optimized graph structure is encoded using GNN, and structural embedding captures the topological structure information of nodes in the graph. The former can help the model understand the deep semantic relationships between node attributes, especially when the attribute text is complex and diverse, while the latter can help the model capture the connection patterns between nodes and the global characteristics of the graph structure.

[0049] First, hierarchical semantic encoding is performed on the nodes of the marine oil and gas system. For equipment operating at a specific time, local time windows are extracted. Equipment logs are then divided into blocks according to fixed time windows, and a lightweight LLM (DistilBERT is used in this paper) is input to obtain local semantic representations. , A sliding window strategy (step size = 1) is used to cover all time series data.

[0050] For continuously operating devices, perform global temporal feature extraction. Input the entire device historical log (24 hours) into the LLM and extract the global context representation: , in To enhance the weight of fault-related words using TF-IDF-based keyword attention masks.

[0051] Design dynamic gating that integrates local and global features: , For the Sigmoid function, These are learnable parameters.

[0052] S4.2 Structural Embedding Learning (GNN Side) To adapt to the needs of marine scenarios, this paper uses a spatiotemporal graph convolutional network (ST-GNN) to acquire information at both 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 relationships of the underwater production system. , Based on ocean characteristic parameters, in the attention coefficient Introducing environmental disturbance factors, the results are calculated using node similarity: , in The edge weights are determined by comprehensively considering signal communication quality (SNR≥20dB). η =1.0, otherwise linear decay) and ocean current impact intensity (ADCP measured current velocity >1m / s) η ← η (×0.8) Two key parameters.

[0054] To capture long-term temporal patterns in deep-water equipment signals, dilated causal convolution (DilatedTCN) is used. , in l For the number of floors, k This represents the kernel size.

[0055] Perform residual joins and hierarchical 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 (topology vs. semantic description) of equipment nodes. Topological features (GNN side): reflect the physical connections (e.g., umbilical cable communication, hydraulic control links) and dynamic interactions (e.g., sensor data drift under ocean current disturbances) between underwater equipment. Semantic features (LLM side): include unstructured text descriptions such as maintenance logs (e.g., "underwater valve seal failure") and ROV inspection reports (e.g., "abnormal anode block detachment"). Discriminator D It is necessary to distinguish the source of embedding (semantic / topological) and adapt to dynamic disturbances in the marine environment.

[0057] Considering the above factors, a neighborhood discriminator is designed and constructed. D Distinguish between embedding sources (semantic / structural): , Adversarial training objective, step 1: Discriminator loss (maximizing source classification accuracy): , Step 2: Generator Loss (Minimizing Modal Difference): , Compared to traditional GRL, which multiplies the discriminator gradient during backpropagation... λ =−1, in dynamic marine environments, adaptive weighting needs to be introduced: , Step S5 specifically includes: S5.1: Semantic embedding and structural embedding focus on the semantic attributes and topological structure of nodes, respectively, and there is a certain degree of information isolation between them. To fully utilize the complementarity of these two embeddings, a multi-head attention mechanism is used to fuse them. The multi-head attention mechanism can dynamically weigh the importance of semantic and structural information and adaptively adjust the contributions of the two embeddings according to the needs of specific tasks, thereby generating a more comprehensive and expressive node representation.

[0058] S5.2: We perform linear transformations on semantic embedding and structural embedding respectively to generate query vector, key vector, and value vector: , Where Qs, Ks, and Vs represent the query vector, key vector, and value vector of the semantic embedding, respectively. These represent the query vector, key vector, and value vector of the structure embedding, respectively. These represent the weight matrices of the linear transformation.

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

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

[0061] Step S6 specifically includes: S6.1: Residual connections are used to fuse semantic information back into the original features, avoiding information loss caused by direct overwriting. , in, It is a learnable linear transformation matrix used for aligning dimensions.

[0062] S6.2: Based on the fusion embedding to execute downstream tasks (such as equipment fault prediction and network anomaly detection, depending on the specific task requirements, which is not the focus of this article), the network parameters are jointly optimized 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, GCN is first used to aggregate node neighbor information through a normalized adjacency matrix to obtain the node embedding H: , in For degree matrix, These are trainable parameters. Then, the failure probability of the marine oil and gas equipment is obtained through mapping: , The output range [0,1] represents the probability of the device malfunctioning.

[0064] Example 2 This embodiment provides a marine oil and gas equipment data monitoring system based on augmented graph learning, including: The data acquisition module is configured as follows: A computer-readable storage medium storing a plurality of instructions adapted for loading and execution by a processor of a terminal device of the aforementioned method for monitoring marine oil and gas equipment data based on augmented graph learning.

[0065] A terminal device includes a processor and a computer-readable storage medium, the processor being used to implement various instructions; the computer-readable storage medium being used to store multiple instructions, the instructions being adapted to be loaded and executed by the processor as described in the method for monitoring marine oil and gas equipment data based on augmented 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, all equivalent changes made in accordance with the structure, shape and principle of the present invention should be covered within the scope of protection of the present invention.

Claims

1. A data monitoring method for marine oil and gas equipment based on reinforcement graph learning, characterized in that, include: Acquire multi-source sensor data and construct a dynamic network structure graph G=(V,A,X,Y) with spatiotemporal characteristics; where the node set V={v_1,v_2,...,v_N} represents various device entities, and each node v_i contains the device ID, type, spatial coordinates, and installation location of the device; the adjacency matrix A accurately characterizes the complex relationships between nodes through three dimensions, including hard relationships based on physical connections, soft relationships based on data correlation, and semantic relationships based on maintenance records; the node feature set X={x_1,x_2,...,x_N} is a multi-dimensional spatiotemporal tensor, and each x_i integrates the D-dimensional features of the device over T time steps, including numerical sensor readings, textual maintenance logs, and inherent device attributes; the device label set Y is flexibly defined according to the specific monitoring task, including discrete fault type codes and continuous remaining life prediction values; The adjacency matrix A is upgraded from a static definition to a spatiotemporal correlation matrix, and the feature matrix X integrates four types of modal features, represented as: ， Where i and j are node numbers. The parameters are learnable, and the contribution of each mode is dynamically adjusted through backpropagation; physical connection weights. Based on device communication links; semantic association weights LLM is used to parse device logs to generate relationship descriptions, and cosine similarity is calculated to reveal implicit associations between devices; causal time-series weighting is employed. Alignment of Granger causality test with DTW time series; preprocessing and text-based transformation of the original node attributes of the dynamic network structure graph; The transformed dynamic network structure graph is analyzed using a Large Language Model (LLM). This includes analyzing node attribute text using LLM to infer potential relationships between nodes. For each pair of nodes, its attribute text is input into the LLM, and the relationship description output by the LLM is obtained. The relationship weight between each pair of nodes is calculated based on the similarity of the attribute text vectors between nodes, as shown below: , in, This represents the weight of the relationship between node i and node j. , and Let i and j represent the attribute text vectors of node i and node j, respectively. This represents the similarity between the attribute text vectors of node i and node j. This represents the similarity between the attribute text vectors of node i and node k, where n represents the total number of nodes and k is the number of nodes. The method of using a Large Language Model (LLM) to analyze the transformed dynamic network structure graph also includes extracting the topological features of underwater equipment nodes using a graph computation method based on adjacency matrices. Specifically, for deep-sea production networks, the degree centrality of each node is calculated as a basic structural attribute, and a dynamically weighted PageRank algorithm is used to accurately evaluate the global importance of underwater equipment nodes. This is achieved by introducing time-varying communication weights and environmental disturbance factors to adapt to complex deep-sea conditions. The iterative calculation process is expressed as follows: ， in, This indicates the nodes constructed by the marine oil and gas equipment system. vi The neighborhood group, This indicates the nodes constructed by the marine oil and gas equipment system. The neighborhood group, N The total number of all built device nodes. These are time-varying edge weights, reflecting the node's... vj and vk At any moment t Real-time communication load, It is an environmental disturbance factor, damping coefficient Set it to 0.8; Feature extraction is performed from both the semantic embedding side and the structural embedding side. Utilizing a multi-head attention mechanism to fuse semantic embedding and structural embedding features; Data monitoring is achieved by embedding downstream tasks into the system.

2. The method for monitoring marine oil and gas equipment data based on augmented graph learning according to claim 1, characterized in that, The process of acquiring multi-source sensor data and constructing a dynamic network structure graph with spatiotemporal characteristics includes expanding the basic node set V, upgrading each device node vi∈V to an enhanced node entity with spatiotemporal awareness, 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.

3. The method for monitoring marine oil and gas equipment data based on augmented graph learning according to claim 2, characterized in that, The preprocessing and text-based transformation of the original node attributes of the dynamic network structure graph includes: for continuous attributes, mapping them 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 fragment for each node, which contains all attribute information of that node, and sorting and describing the attributes according to the weights of semantic associations, as follows: The attributes of the node [node number] include: [attribute1 name]: [attribute1 value], [attribute2 name]: [attribute2 value], ..., [attribute n name]: [attribute n value]. Among them, the attribute names and attribute values ​​are sorted according to the weight of semantic association, with the attribute with the higher weight appearing first.

4. The method for monitoring marine oil and gas equipment data based on augmented graph learning according to claim 3, characterized in that, The method of using a Large Language Model (LLM) to parse the transformed dynamic network structure graph also includes using LLM for edge adjustment and adaptive graph structure optimization. First, LLM is used to parse node pairs (…). vi , vj The text description of the relation is output, along with the relation type and confidence level, and a time-series relation tensor is constructed. Historical information is aggregated using LSTM, and then topological and semantic features are dynamically fused, represented as follows: ， Among them, the balance factor β Learning jointly from topological and semantic features, 。 5. The method for monitoring marine oil and gas equipment data based on augmented graph learning according to claim 4, characterized in that, The feature extraction is performed from both the semantic embedding side and the structural embedding side. On the semantic embedding side, hierarchical semantic encoding is performed on the construction nodes of the marine oil and gas system. For equipment operating at a specific time, local time windows are divided for feature extraction. Equipment logs are divided into blocks according to fixed time windows, and a lightweight LLM is input to obtain local semantic representations. A sliding window strategy is used to cover all time-series data; for continuously operating equipment, global time-series feature extraction is performed. Input the entire device history log into the LLM and extract the global context representation. The design of dynamic gating, which integrates local and global features, is represented as follows: ， in, For the Sigmoid function, For learnable parameters, It is a bias term.

6. The method for monitoring marine oil and gas equipment data based on augmented graph learning according to claim 5, characterized in that, The feature extraction process, performed from both semantic and structural embedding sides, further includes acquiring temporal and spatial information using a spatiotemporal graph convolutional network on the structural embedding side. First, spatial dependency modeling is performed, employing a multi-scale graph attention network to aggregate the topological relationships of the underwater production system. Based on oceanographic parameters, an environmental disturbance factor is introduced into the attention coefficient, and the probability of connections between nodes is calculated using node similarity. For the long-period characteristics of deep-water equipment signals, dilated causal convolution is used to capture long-term temporal patterns, followed by residual connections and hierarchical normalization. Node similarity is expressed as: ， in The edge weight is denoted as .

7. The method for monitoring marine oil and gas equipment data based on augmented graph learning according to claim 6, characterized in that, The method of using a multi-head attention mechanism to fuse semantic and structural embedding features includes identifying and distinguishing embedding sources based on adversarial cross-modal alignment, performing linear transformations on semantic and structural embeddings respectively to generate query vectors, key vectors, and value vectors, calculating attention weights between semantic and structural embeddings to measure their importance in the fusion process, and using residual connections to fuse semantic information back to the original features to avoid information loss due to direct overwriting. Finally, downstream tasks are executed based on the fused embeddings, and the 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.

8. A marine oil and gas equipment data monitoring system based on augmented graph learning, executing the marine oil and gas equipment data monitoring method based on augmented graph learning as described in claim 1, characterized in that, include: The data acquisition module is configured to acquire multi-source sensor data and construct a dynamic network structure diagram with spatiotemporal characteristics. The preprocessing module is configured to preprocess and convert the raw node attributes of the dynamic network structure graph into text. The parsing module is configured to parse the transformed dynamic network structure graph using the Large Language Model (LLM). The feature extraction module is configured to extract features from both the semantic embedding side and the structural embedding side. The feature fusion module is configured to perform semantic embedding and structural embedding feature fusion using a multi-head attention mechanism; The monitoring module is configured to perform downstream tasks based on fusion embedding to achieve data monitoring.

Images