Digital twin drive multi-source heterogeneous data fusion and intelligent analysis method for oil and gas pipeline construction

Abstract

The invention discloses a digital twin-driven multi-source heterogeneous data fusion and intelligent analysis method for oil and gas pipeline construction. The method comprises the following steps: collecting multi-source heterogeneous data related to oil and gas pipeline construction; fusing data by using a space-time diagram neural network and constructing a dynamic offset compensation model to correct coordinate drift; constructing a data credibility dynamic evaluation chain, performing cross validation on sensor data to generate a confidence coefficient weight, and complementing missing data through a GAN complementing module to form a credibility enhanced data set; designing a layered distillation compression algorithm to compress the PINN model into a TinyLSTM engine, and developing a model increment updating protocol to realize efficient collaboration of a cloud end and an edge end; and training and optimizing a TinyLSTM engine by using the data set with enhanced credibility. According to the method, the defects in the aspects of data dynamic correction, credibility evaluation and complementation, model edge deployment real-time response and the like in the prior art are overcome, and the method has remarkable innovativeness and practicability.

Description

Technical Field

[0001] The present invention relates to the field of oil and gas pipeline technology, and in particular to a digital twin-driven multi-source heterogeneous data fusion and intelligent analysis method for oil and gas pipeline construction. Background Art

[0002] The construction of oil and gas pipelines requires processing and analyzing massive amounts of heterogeneous, multi-source data, including time-series data from pipeline strain sensors and InSAR satellite remote sensing surface deformation data. However, existing technologies present the following challenges: First, it is difficult to dynamically correct for coordinate drift caused by geological subsidence and pipeline displacement, affecting data accuracy; second, they rely on preset thresholds to identify fault signs, making them susceptible to false alarms due to sensor failure or environmental noise; and third, digital twins rely on cloud computing power, unable to meet the millisecond-level response requirements for micro-leaks and lacking real-time performance.

[0003] These problems restrict the intelligent development of oil and gas pipeline construction. There is an urgent need for a method that can effectively integrate multi-source heterogeneous data, dynamically correct data drift, reduce false alarms, and have edge computing capabilities to improve the efficiency and safety of oil and gas pipeline construction. Summary of the Invention

[0004] The purpose of this invention is to solve the shortcomings of the existing technology and propose a digital twin-driven multi-source heterogeneous data fusion and intelligent analysis method for oil and gas pipeline construction.

[0005] In order to achieve the above object, the present invention adopts the following technical solutions:

[0006] The digital twin-driven multi-source heterogeneous data fusion and intelligent analysis method for oil and gas pipeline construction includes the following specific steps:

[0007] S1: Multi-source heterogeneous data acquisition: Collect multi-source heterogeneous data related to oil and gas pipeline construction, including pipeline strain sensor time series data, InSAR satellite surface deformation data, GIS geographic information, GPS positioning data and surrounding environment data, to build a comprehensive monitoring network;

[0008] S2: Data fusion and dynamic compensation: Pipeline strain data and InSAR surface deformation data are fused through a spatiotemporal graph neural network (STGNN), temporal and spatial correlation features are learned, and a dynamic offset compensation model is constructed. This model is used to correct coordinate drift in real time and output corrected 3D coordinates and strain data.

[0009] S3: Data credibility assessment and completion: Distributed nodes cross-validate sensor data to generate dynamic confidence weights. For missing data in extreme working conditions, physical models and historical data are used to generate synthetic data, and confidence weights are combined to generate a credible data set.

[0010] S4: Edge Intelligent Optimization: A layered distillation algorithm is used to compress the cloud-based PINN model into a lightweight TinyLSTM engine, adapted for real-time computing on edge devices. Local training on edge devices generates incremental updates, which are aggregated in the cloud and then distributed to reduce bandwidth usage and achieve continuous model optimization.

[0011] S5: Decision support application: Based on the optimized model and data above, the TinyLSTM engine is trained and optimized using a credibility-enhanced dataset, enabling it to perform risk assessment, construction progress forecasting, resource optimization and design optimization during oil and gas pipeline construction based on real-time monitoring data, providing decision support for oil and gas pipeline construction.

[0012] As a further technical solution of the present invention, S2 specifically includes:

[0013] S21: Construct a spatiotemporal graph neural network (STGNN) model. This model takes the collected pipeline strain sensor time series data and InSAR satellite remote sensing surface deformation data as input. The STGNN model fuses the two data sources, learns the feature associations of the data in the spatiotemporal dimension, and constructs a dynamic offset compensation model.

[0014] S22: Using the dynamic offset compensation model, the coordinate data of the pipeline strain sensor is corrected in real time to compensate for the coordinate drift caused by geological settlement and pipeline displacement, and the corrected pipeline three-dimensional coordinate information and strain data are output.

[0015] As a further technical solution of the present invention, the S21 specifically includes:

[0016] S211: Data preprocessing and fusion:

[0017] The collected pipeline strain sensor time series data is S(t)={s1(t),s2(t),...,s n (t)}, n represents the number of sensors, t represents the time series; InSAR satellite remote sensing surface deformation data is I(t)={i1(t),i2(t),...,i m (t)}, m represents the number of InSAR observation points, and t represents the time series;

[0018] The two data are aligned in the time dimension to ensure data consistency. The missing time points are filled by interpolation method so that the two data have the same time step. The aligned pipeline strain sensor time series data and InSAR satellite remote sensing surface deformation data are fused to construct the fused data matrix X(t) = [S(t); I(t)], where the semicolon represents the row splicing of the matrix;

[0019] S212: Space-time graph construction:

[0020] Define the graph structure G = (V, E), where V represents the node set, including pipeline strain sensors and InSAR observation points, and their location relationship in geographic space; E represents the edge set between nodes, reflecting the spatial correlation between nodes;

[0021] Calculate the distance matrix D between nodes, determine the connection relationship between nodes based on the distance threshold, and establish an edge between two nodes if the distance between them is less than the threshold;

[0022] Construct the adjacency matrix A of the graph, where the element A ij Represents the connection weight between node i and node j. The weight can be calculated based on the distance between nodes or other correlation metrics;

[0023] S213: Spatiotemporal graph convolution operation:

[0024] In the spatiotemporal graph neural network (STGNN), graph convolution operations are used to capture the spatial feature associations between nodes;

[0025] For capturing feature correlation in the time dimension, a one-dimensional convolution operation is used;

[0026] S214: Dynamic offset compensation model construction: In the output layer of STGNN, the time attenuation coefficient α and the spatial offset coefficient β are introduced to construct a dynamic offset compensation model.

[0027] As a further technical solution of the present invention, the S22 specifically includes:

[0028] S221: Define coordinate data and strain data: Assume that the original three-dimensional coordinate data of the pipeline strain sensor is P(t) = [x(t), y(t), z(t)] T , where x(t), y(t), and z(t) represent the east, north, and height coordinate components of the sensor at time t, respectively, and T represents the matrix transpose; the original strain data is Q(t) = {ε axial (t),ε hoop (t)}, where ε axial (t) represents the axial strain, ε hoop (t) represents hoop strain;

[0029] S222: Obtaining an offset compensation amount: Obtaining an offset compensation amount ΔP(t) according to the dynamic offset compensation model in step S214;

[0030] S223: Coordinate data correction: Corrected pipeline three-dimensional coordinate information P corrected (t) = P(t) + ΔP(t);

[0031] S224: Strain data correction: Considering the influence of coordinate offset on pipeline strain, the strain correction coefficient k is introduced. ε , which is defined as: Where: ||ΔP(t)|| represents the Euclidean norm of the offset compensation ΔP(t), L0 is the original length of the pipeline; the corrected strain data

[0032] S225: Model training and parameter update: Define the loss function L for training the dynamic offset compensation model, combining the coordinate correction error and the strain correction error; use an optimization algorithm (such as Adam) to update the model parameters α and β to minimize the loss function L and improve the model's correction accuracy for coordinate drift and strain data.

[0033] As a further technical solution of the present invention, S3 specifically includes:

[0034] S31: Build a dynamic data credibility evaluation chain. Based on the federated learning framework, cross-validate sensor data distributed across multiple nodes and generate confidence weights for each sensor data.

[0035] S32: Develop a generative adversarial network (GAN) completion module. For missing pipeline data under extreme working conditions, use the existing pipeline physical model and historical data to generate synthetic data that conforms to physical laws through the GAN module to complete the missing data. The completed data is weighted in combination with confidence weights to generate a dataset with enhanced credibility.

[0036] As a further technical solution of the present invention, the S31 specifically includes:

[0037] S311: Define sensor data: Assume that the sensor data distributed on multiple nodes is D = {D1, D2, ..., D m}, where m represents the number of sensor nodes, D i ={d i1 ,d i2 ,...,d in} represents the data collected by the i-th node at n time steps; each data point d ij Contains sensor measurements and timestamps, i.e., d ij =(v ij ,t j ), v ij is the measured value, t j is the corresponding time;

[0038] S312: Initialize the federated learning framework: In the federated learning framework, set up a central server and multiple clients (one for each sensor node), define the global model parameter θ, initialize it to θ0, and set the learning rate α, momentum coefficient β, and model decay coefficient γ.

[0039] S313: Data consistency evaluation: For each sensor node i, calculate the consistency score C of its data with the data of other nodes i ;

[0040] S314: Data stability evaluation: Calculate the stability score S of each sensor node i data i , based on the time series variation characteristics of data;

[0041] S315: Confidence weight calculation: Combine the consistency and stability scores to generate the confidence weight W of each sensor data i , fused through quadratic functions:

[0042] S316: Federated Learning Model Update:

[0043] Each client uses local data D i and the current global model parameters θ, calculate the gradient

[0044] And use the gradient descent method with momentum to update the local model parameters: in: is the gradient value of the previous round, initially 0;

[0045] The client will update the model parameters θ i ′ is sent to the central server, which aggregates the parameters of all clients and updates the global model parameters:

[0046] As a further technical solution of the present invention, the S32 specifically includes:

[0047] S321: Define the GAN module structure: The GAN module consists of a generator G and a discriminator D. The input of the generator G is a random noise vector z and some known pipeline data X. known , the output is the completed pipeline data X generated , the input of the discriminator D is the real pipeline data X real Or the data X generated by the generator generated , the output is the judgment result P real, represents the probability that the input data is true data;

[0048] S322: Combining pipeline physical models and historical data: Define a pipeline physical model M, which combines pipeline data with physical constraints based on physical laws such as pipeline material, structure, and mechanical properties. Collect historical data H, including pipeline data collected in the past under similar operating conditions, to train the GAN module and help it learn the distribution characteristics of real data.

[0049] S323: Generator training: The goal of the generator G is to minimize the difference between the generated data and the real data while satisfying the physical model constraints. Its loss function L G Defined as: in: represents the expectation; λ is the weight coefficient of the physical constraint loss, which is used to balance the authenticity and physical rationality of the generated data;

[0050] S324: Discriminator training: The goal of the discriminator D is to distinguish between real data and generated data, and its loss function L D Defined as:

[0051] S325: Generate synthetic data: After training is completed, for the missing pipeline data X under extreme working conditions missing , use the generator G to generate the complete dataset X complete ;

[0052] S326: Combined with confidence weight weighting processing: For the completed data X complete , combined with the confidence weight W obtained in step S315 i , generate a credibility-enhanced dataset X enhanced : Where: X complete,i represents the completion data generated by the generator for the i-th sensor node.

[0053] As a further technical solution of the present invention, the S4 specifically includes:

[0054] S41: Design a hierarchical distillation compression algorithm to compress the physical information neural network (PINN) model trained in the cloud into a TinyLSTM engine suitable for edge device deployment, enabling the edge device to run the TinyLSTM engine in real time for pipeline deformation and buckling calculations.

[0055] S42: Develop a model incremental update protocol. The edge node uses the newly collected local data to locally train the TinyLSTM engine, and sends the incremental model update information obtained from the local training to the cloud. The cloud aggregates and iterates the incremental update information of multiple edge nodes to generate a new model update package, and then sends the update package to the edge node to realize the incremental update of the model on the edge device and reduce the data transmission bandwidth usage.

[0056] As a further technical solution of the present invention, the S41 specifically includes:

[0057] S411: Define the structure of the layered distillation compression algorithm:

[0058] Cloud-based Physical Information Neural Network (PINN) model: The PINN model is a deep neural network whose input is pipeline physical parameters (such as pressure, temperature, and strain) and outputs pipeline deformation and buckling predictions. The model structure includes multiple fully connected layers, activation function layers, and physical constraint layers.

[0059] Edge TinyLSTM Engine: The TinyLSTM engine is a lightweight recurrent neural network (RNN) structure designed specifically for edge devices. Its input is a time series of pipeline physical parameters, and its output is a real-time pipeline deformation and buckling prediction value.

[0060] S412: Layered Distillation Compression Algorithm:

[0061] Knowledge distillation loss function: Define the knowledge distillation loss function L distill , used to measure the difference between the output of the PINN model and the TinyLSTM engine;

[0062] Hierarchical distillation process: the output Y of the PINN model PINN As the output of the teacher model, the output Y of the TinyLSTM engine is TinyLSTM As the output of the student model, during the training process, the knowledge distillation loss function L is minimized. distill , so that the TinyLSTM engine can learn the knowledge of the PINN model;

[0063] Model compression steps: Prune the PINN model weights to remove unimportant weights and reduce the number of model parameters. The pruning criteria can be based on the absolute value of the weights, retaining larger weights. Quantize the PINN model weights from floating point numbers to low-precision integers to reduce model storage space. Use the knowledge distillation method, using the output of the pruned and quantized PINN model as a guide to train the TinyLSTM engine.

[0064] S413: Edge Device Deployment:

[0065] TinyLSTM engine optimization: Further optimization of the trained TinyLSTM engine, such as model pruning and quantization, is performed to adapt to the computing and storage limitations of edge devices.

[0066] Real-time deformation and buckling calculation: The optimized TinyLSTM engine is deployed on edge devices. The edge devices collect real-time pipeline physical parameters and input them into the TinyLSTM engine for real-time pipeline deformation and buckling calculation.

[0067] As a further technical solution of the present invention, the S42 specifically includes:

[0068] S421: Local training of edge nodes: defining local training data, updating local models, calculating local gradients, and generating incremental update information;

[0069] S422: Cloud aggregation iteration: collect incremental update information, define aggregation weights, aggregate incremental update information, and update global model parameters;

[0070] S423: Model update distribution: The cloud distributes the updated global model parameters Packaged into model update package P update , the cloud will P update Send it to each edge node. After receiving the update package, the edge node replaces the local model parameters with Complete the incremental update of the model.

[0071] The beneficial effects of the present invention are:

[0072] 1. Data fusion and dynamic correction: The spatial-temporal graph neural network (STGNN) is introduced to fuse pipeline strain sensor time series data with InSAR satellite remote sensing surface deformation data. A dynamic offset compensation model is constructed to correct coordinate drift caused by geological subsidence and pipeline displacement in real time. This significantly improves data accuracy and reliability, providing a solid data foundation for subsequent analysis and decision-making.

[0073] 2. Data credibility assessment and completion: Build a dynamic data credibility assessment chain, cross-validate multi-node sensor data based on a federated learning framework, generate confidence weights, and develop a generative adversarial network (GAN) completion module to generate synthetic data that conforms to physical laws for missing data in extreme working conditions, reducing the false alarm rate and enhancing the system's monitoring and fault diagnosis capabilities in complex environments.

[0074] 3. Efficient model deployment and update: We designed a layered distillation compression algorithm to compress the physical information neural network (PINN) model trained in the cloud into a TinyLSTM engine suitable for edge device deployment. We also developed an incremental model update protocol to achieve efficient collaboration between the cloud and the edge. This reduces data transmission bandwidth usage, improves the system's real-time performance and response speed, and meets the real-time monitoring and intelligent analysis needs of oil and gas pipeline construction sites.

[0075] 4. Intelligent Decision Support: Based on optimized models and credibility-enhanced datasets, the TinyLSTM engine is used to conduct risk assessment, construction progress forecasting, resource optimization and allocation, and design optimization during oil and gas pipeline construction. This provides scientific decision support for oil and gas pipeline construction, helps reduce construction costs, improve construction safety and reliability, and has significant economic and social benefits. BRIEF DESCRIPTION OF THE DRAWINGS

[0076] Figure 1 This is a flow chart of the digital twin-driven multi-source heterogeneous data fusion and intelligent analysis method for oil and gas pipeline construction proposed in this invention. DETAILED DESCRIPTION

[0077] In order to make the technical means, creative features, objectives and effects achieved by the present invention easier to understand, the present invention is further described below in conjunction with specific implementation methods.

[0078] Please see the attached Figure 1 The digital twin-driven multi-source heterogeneous data fusion and intelligent analysis method for oil and gas pipeline construction includes the following specific steps:

[0079] S1: Multi-source heterogeneous data acquisition: Collect multi-source heterogeneous data related to oil and gas pipeline construction, including pipeline strain sensor time series data, InSAR satellite surface deformation data, GIS geographic information, GPS positioning data and surrounding environment data, to build a comprehensive monitoring network;

[0080] S2: Data fusion and dynamic compensation: Pipeline strain data and InSAR surface deformation data are fused through a spatiotemporal graph neural network (STGNN), temporal and spatial correlation features are learned, and a dynamic offset compensation model is constructed. This model is used to correct coordinate drift in real time and output corrected 3D coordinates and strain data.

[0081] S21: Construct a spatiotemporal graph neural network (STGNN) model. This model takes the collected pipeline strain sensor time series data and InSAR satellite remote sensing surface deformation data as input. The STGNN model fuses the two data sources, learns the feature associations of the data in the spatiotemporal dimension, and constructs a dynamic offset compensation model.

[0082] S211: Data preprocessing and fusion:

[0083] The collected pipeline strain sensor time series data is S(t)={s1(t),s2(t),...,s n (t)}, n represents the number of sensors, t represents the time series; InSAR satellite remote sensing surface deformation data is I(t)={i1(t),i2(t),...,i m (t)}, m represents the number of InSAR observation points, and t represents the time series;

[0084] The two data are aligned in the time dimension to ensure data consistency. The missing time points are filled by interpolation method so that the two data have the same time step. The aligned pipeline strain sensor time series data and InSAR satellite remote sensing surface deformation data are fused to construct the fused data matrix X(t) = [S(t); I(t)], where the semicolon represents the row splicing of the matrix;

[0085] S212: Space-time graph construction:

[0086] Define the graph structure G = (V, E), where V represents the node set, including pipeline strain sensors and InSAR observation points, and their location relationship in geographic space; E represents the edge set between nodes, reflecting the spatial correlation between nodes;

[0087] Calculate the distance matrix D between nodes, determine the connection relationship between nodes based on the distance threshold, and establish an edge between two nodes if the distance between them is less than the threshold;

[0088] Construct the adjacency matrix A of the graph, where the element A ij Represents the connection weight between node i and node j. The weight can be calculated based on the distance between nodes or other correlation metrics;

[0089] S213: Spatiotemporal graph convolution operation:

[0090] In the spatiotemporal graph neural network (STGNN), graph convolution operations are used to capture the spatial feature associations between nodes. The graph convolution operation can be expressed as: Among them: H (k) Represents the node feature matrix of the kth layer; A+I represents the adjacency matrix after adding self-connection, and I is the identity matrix; yes The degree matrix of W (k)is the trainable weight matrix of the kth layer; σ represents the activation function, such as ReLU;

[0091] For feature association capture in the time dimension, a one-dimensional convolution operation is used. Assuming that the time step is T and the number of input feature channels is F in , the number of output feature channels is F out , then the one-dimensional convolution operation can be expressed as: Where: X t represents the input feature vector at time step t; K represents the size of the convolution kernel; W τ represents the weight matrix of the τth convolution kernel; b represents the bias term; Y t Represents the output feature vector after the convolution operation;

[0092] S214: Dynamic offset compensation model construction: In the output layer of STGNN, the time attenuation coefficient α and the spatial offset coefficient β are introduced to construct a dynamic offset compensation model, which can be expressed as: Where: ΔP(t) represents the coordinate offset compensation at time step t; α is the time attenuation coefficient, which is used to control the influence weight of feature changes in the time dimension on offset compensation, and its value range is [0, 1]. β is the spatial offset coefficient, which is used to control the influence weight of offset changes in the spatial dimension, and its value range is [0, 1]. STGNN(X(t)) represents the output feature vector of the STGNN model for the fused data X(t). Represents the coordinate gradient at time step t, which can be obtained by differential calculation of the coordinate data.

[0093] S22: Using a dynamic offset compensation model, the coordinate data of the pipeline strain sensor is corrected in real time to compensate for coordinate drift caused by geological settlement and pipeline displacement, and the corrected pipeline three-dimensional coordinate information and strain data are output;

[0094] S221: Define coordinate data and strain data: Assume that the original three-dimensional coordinate data of the pipeline strain sensor is P(t) = [x(t), y(t), z(t)] T , where x(t), y(t), and z(t) represent the east, north, and height coordinate components of the sensor at time t, respectively, and T represents the matrix transpose; the original strain data is Q(t) = {ε axial (t),ε hoop (t)}, where ε axial (t) represents the axial strain, ε hoop (t) represents hoop strain;

[0095] S222: Obtaining an offset compensation amount: Obtaining an offset compensation amount ΔP(t) according to the dynamic offset compensation model in step S214;

[0096] S223: Coordinate data correction: Corrected pipeline three-dimensional coordinate information P corrected (t) = P(t) + ΔP(t);

[0097] S224: Strain data correction: Considering the influence of coordinate offset on pipeline strain, the strain correction coefficient k is introduced. ε , which is defined as: Where: ||ΔP(t)|| represents the Euclidean norm of the offset compensation ΔP(t), L0 is the original length of the pipeline; the corrected strain data

[0098] S225: Model training and parameter update: Define the loss function L for training the dynamic offset compensation model, combining the coordinate correction error and the strain correction error: Where: P true (t) is the real coordinate data obtained by high-precision measuring equipment; S true (t) is the real strain data; T is the time step number of training data; λ1, λ2 are weight coefficients, which control the proportion of coordinate correction error and strain correction error in the total loss, generally satisfying λ1+λ2=1; an optimization algorithm (such as Adam) is used to update the model parameters α and β to minimize the loss function L and improve the model's correction accuracy for coordinate drift and strain data.

[0099] Step S2 constructs a spatiotemporal graph neural network (STGNN) model, integrates pipeline strain sensor time series data and InSAR satellite remote sensing surface deformation data, learns the feature association of data in the spatiotemporal dimensions, and uses a dynamic offset compensation model to perform real-time correction of pipeline coordinate data. This solves the problem in existing technologies that it is difficult to dynamically correct coordinate drift caused by geological subsidence and pipeline displacement, and improves the accuracy and reliability of the data.

[0100] S3: Data credibility assessment and completion: Distributed nodes cross-validate sensor data to generate dynamic confidence weights. For missing data in extreme working conditions, physical models and historical data are used to generate synthetic data, and confidence weights are combined to generate a credible data set.

[0101] S31: Build a dynamic data credibility evaluation chain. Based on the federated learning framework, cross-validate sensor data distributed across multiple nodes and generate confidence weights for each sensor data.

[0102] S311: Define sensor data: Assume that the sensor data distributed on multiple nodes is D = {D1, D2, ..., D m}, where m represents the number of sensor nodes, D i ={d i1 ,d i2 ,...,d in} represents the data collected by the i-th node at n time steps; each data point d ij Contains sensor measurements and timestamps, i.e., d ij =(v ij ,t j ), v ij is the measured value, t j is the corresponding time;

[0103] S312: Initialization of the federated learning framework: In the federated learning framework, a central server and multiple clients are set up (one for each sensor node). The global model parameter θ is defined and initialized to θ0. This model is used to evaluate the consistency and stability of the data. The learning rate α, momentum coefficient β, and model decay coefficient γ are set, where α∈(0,1), β∈[0,1), and γ∈(0,1). These are used to control parameter updates during model training.

[0104] S313: Data consistency evaluation: For each sensor node i, calculate the consistency score C of its data with the data of other nodes i , the formula is as follows: Where: σ is the bandwidth parameter of the consistency Gaussian kernel, which is used to control the impact of data differences on the consistency score; || D i -D k || is the Euclidean distance between the data of node i and node k;

[0105] S314: Data stability evaluation: Calculate the stability score S of each sensor node i data i , based on the time series change characteristics of the data: Where: II is the indicator function, which takes 1 when the condition is met and 0 otherwise; τ is the stability threshold, which is used to determine whether the data change is within a reasonable range. If the data change amplitude of adjacent time steps is less than τ, the data is considered to be stable at that time step;

[0106] S315: Confidence weight calculation: Combine the consistency and stability scores to generate the confidence weight W of each sensor data i , fused through quadratic functions:

[0107] S316: Federated Learning Model Update:

[0108] Each client uses local data D i and the current global model parameters θ, calculate the gradient

[0109] S3161: Define local loss function: The client is based on local data D i And model parameters θ, define the loss function Used to measure the prediction error of the model on local data. For regression problems, the mean square error (MSE) loss function is often used: Where: f θ (x ij ) indicates that the model is input x ij The predicted value under ij is the corresponding true label, n is the number of local data samples;

[0110] S3162: Computing Gradients: Loss Function Find the gradient of the model parameter θ and get the gradient vector Taking a single sample as an example, the gradient calculation formula is: in It represents the gradient of the model prediction value with respect to the parameter θ. Its specific form depends on the model structure used. For example, for a simple linear regression model f θ (x) = θ T x+b, its gradient is: For more complex models (such as neural networks), the gradient needs to be calculated through the backpropagation algorithm;

[0111] And use the gradient descent method with momentum to update the local model parameters: in: is the gradient value of the previous round, initially 0;

[0112] The client will update the model parameters θ i ′ is sent to the central server, which aggregates the parameters of all clients and updates the global model parameters:

[0113] In this way, multi-node sensor data is cross-validated through the federated learning framework, and the confidence weight of each sensor data is generated based on the consistency and stability of data from different nodes, providing a reliable data quality assessment basis for subsequent data fusion and analysis, effectively reducing the risk of false alarms caused by sensor failure or environmental noise interference, and improving the robustness and data quality of the entire system.

[0114] S32: Develop a Generative Adversarial Network (GAN) completion module. For missing pipeline data under extreme operating conditions, the module uses existing pipeline physical models and historical data to generate synthetic data that conforms to physical laws to complete the missing data. The completed data is weighted based on confidence weights to generate a dataset with enhanced credibility.

[0115] S321: Define the GAN module structure: The GAN module consists of a generator G and a discriminator D. The input of the generator G is a random noise vector z and some known pipeline data X. known , the output is the completed pipeline data X generated , the input of the discriminator D is the real pipeline data X real Or the data X generated by the generator generated , the output is the judgment result P real , represents the probability that the input data is true data;

[0116] S322: Combining the pipeline physical model and historical data: Define the pipeline physical model M. This model combines pipeline data with physical constraints based on the physical laws of the pipeline's material, structure, and mechanical properties. For example, the following formula can be used for the pipeline stress-strain relationship: σ = E·ε, where σ is the pipeline stress; E is the elastic modulus of the pipeline material; and ε is the pipeline strain. Collect historical data H, including pipeline data collected in the past under similar operating conditions, to train the GAN module and help it learn the distribution characteristics of real data.

[0117] S323: Generator training: The goal of the generator G is to minimize the difference between the generated data and the real data while satisfying the physical model constraints. Its loss function L G Defined as: in: represents the expectation; λ is the weight coefficient of the physical constraint loss, which is used to balance the authenticity and physical rationality of the generated data;

[0118] S324: Discriminator training: The goal of the discriminator D is to distinguish between real data and generated data, and its loss function L D Defined as:

[0119] S325: Generate synthetic data: After training is completed, for the missing pipeline data X under extreme working conditions missing , use the generator G to generate the complete dataset X complete ;

[0120] S3251: Initialize random noise vector: from Gaussian distribution The random noise vector z is obtained by sampling;

[0121] S3252: Generator input construction: z and known pipeline data X known Splicing along the feature dimension: Input G =[z;X known ];

[0122] S3253: Generate preliminary data output: Input G Input into the generator G to obtain the initial generated data X generated ';

[0123] S3254: Physical model correction: Use pipeline physical model M to X generated 'Correction: X generated =C physics (X generated ′,M);

[0124] S3255: Denormalization: X generated ″ Perform denormalization to restore the value range of the original data: X generated =X generated ″×(max(X f )-min(X f ))+min(X f ), where max(X f ) and min(X f ) represent the maximum and minimum values ​​of f pipeline data respectively;

[0125] S3256: Data Fusion: X generated With known pipeline data X known Fusion, generate complete dataset X complete :X complete =X known ·(1-M missing )+X generated ·M missing ;

[0126] S326: Combined with confidence weight weighting processing: For the completed data X complete, combined with the confidence weight W obtained in step S315 i , generate a credibility-enhanced dataset X enhanced : Where: X complete,i represents the completion data generated by the generator for the i-th sensor node.

[0127] Through the development of the above-mentioned GAN module, existing pipeline physical models and historical data are utilized to generate synthetic data that conforms to physical laws. The completed data is weighted in combination with confidence weights to generate a credibility-enhanced dataset. This effectively solves the problem of missing pipeline data under extreme working conditions, improves data integrity and reliability, and provides a more solid data foundation for intelligent analysis of oil and gas pipeline construction.

[0128] S4: Edge Intelligent Optimization: A layered distillation algorithm is used to compress the cloud-based PINN model into a lightweight TinyLSTM engine, adapted for real-time computing on edge devices. Local training on edge devices generates incremental updates, which are aggregated in the cloud and then distributed to reduce bandwidth usage and achieve continuous model optimization.

[0129] S41: Design a hierarchical distillation compression algorithm to compress the physical information neural network (PINN) model trained in the cloud into a TinyLSTM engine suitable for edge device deployment, enabling the edge device to run the TinyLSTM engine in real time for pipeline deformation and buckling calculations.

[0130] S411: Define the structure of the layered distillation compression algorithm:

[0131] Cloud Physical Information Neural Network (PINN) model: The PINN model is a deep neural network whose input is pipeline physical parameters (such as pressure, temperature, and strain) and whose output is pipeline deformation and buckling prediction values. The model structure includes multiple fully connected layers, activation function layers, and physical constraint layers. For example, the model can be expressed as: PINN: X→Y, where X is the input pipeline physical parameters and Y is the output pipeline deformation and buckling prediction values.

[0132] Edge TinyLSTM Engine: The TinyLSTM engine is a lightweight recurrent neural network (RNN) structure designed specifically for edge devices. Its input is a time series of pipeline physical parameters, and its output is a real-time pipeline deformation and buckling prediction value.

[0133] S412: Layered Distillation Compression Algorithm:

[0134] Knowledge distillation loss function: Define the knowledge distillation loss function L distill, used to measure the difference between the output of the PINN model and the TinyLSTM engine, the loss function can be expressed as: L distill =α·L CE +β·L MSE , where: L CE is the cross entropy loss, used for classification tasks; L MSE is the mean square error loss, used for regression tasks; α and β are weight coefficients used to balance the effects of the two losses;

[0135] Hierarchical distillation process: the output Y of the PINN model PINN As the output of the teacher model, the output Y of the TinyLSTM engine is TinyLSTM As the output of the student model, during the training process, the knowledge distillation loss function L is minimized. distill , so that the TinyLSTM engine can learn the knowledge of the PINN model;

[0136] Model compression steps: Prune the PINN model weights to remove unimportant weights and reduce the number of model parameters. The pruning criteria can be based on the absolute value of the weights, retaining larger weights. Quantize the PINN model weights from floating point numbers to low-precision integers to reduce model storage space. Use the knowledge distillation method, using the output of the pruned and quantized PINN model as a guide to train the TinyLSTM engine.

[0137] S413: Edge Device Deployment:

[0138] TinyLSTM engine optimization: Further optimization of the trained TinyLSTM engine, such as model pruning and quantization, is performed to adapt to the computing and storage limitations of edge devices.

[0139] Real-time deformation and buckling calculation: The optimized TinyLSTM engine is deployed on edge devices. The edge devices collect real-time pipeline physical parameters and input them into the TinyLSTM engine for real-time pipeline deformation and buckling calculation.

[0140] S42: Develop a model incremental update protocol. Edge nodes use newly collected local data to locally train the TinyLSTM engine and send the incremental model update information obtained from the local training to the cloud. The cloud aggregates and iterates the incremental update information from multiple edge nodes to generate a new model update package. This update package is then sent to the edge nodes, enabling incremental updates of the model on edge devices and reducing data transmission bandwidth usage.

[0141] S421: Local training of edge nodes:

[0142] S4211: Define local training data: Edge nodes collect newly collected local pipeline data Dlocal ={d1,d2,...,d i}, where d i represents the i-th data sample, which contains the pipeline physical parameters and the corresponding deformation and buckling labels;

[0143] S4212: Local model update: Edge nodes use D local Perform local training on the TinyLSTM engine and update the local model parameters. The goal of local training is to minimize the loss function L local , defined as: in: is the TinyLSTM engine with parameters θ local Next input x i The predicted value of y i is the corresponding true value of deformation buckling;

[0144] S4213: Calculate local gradient: loss function L local Find the gradient and get the local gradient : in is the model prediction value for parameter θ local gradient;

[0145] S4214: Generate incremental update information: Edge nodes will update local gradients As incremental update information U local Send to the cloud:

[0146] S422: Cloud Aggregation Iteration:

[0147] S4221: Collect incremental update information: The cloud collects incremental update information U1, U2, ..., U from multiple edge nodes. m , where m is the number of edge nodes;

[0148] S4222: Define aggregation weights: Define aggregation weights w1, w2, ..., w based on the amount of data or confidence weight of each edge node. m ,satisfy For example, you can assign weights to nodes based on the number of data samples they have: where n i is the number of data samples of the i-th edge node;

[0149] S4223: Aggregate incremental update information: The cloud performs weighted averaging on the collected incremental update information to generate a global gradient update. :

[0150] S4224: Update global model parameters: Update using global gradients in the cloud Update global model parameters θ global : Where α is the learning rate, which controls the step size of parameter updates;

[0151] S423: Model update distribution: The cloud distributes the updated global model parameters Packaged into model update package P update , the cloud will P update Send it to each edge node. After receiving the update package, the edge node replaces the local model parameters with Complete the incremental update of the model.

[0152] Through the above-mentioned model incremental update protocol, the edge node uses the newly collected local data to locally train the TinyLSTM engine. The cloud aggregates and iterates the incremental update information of multiple edge nodes, generates a new model update package and sends it to the edge node, realizing the incremental update of the model on the edge device and effectively reducing the data transmission bandwidth usage.

[0153] S5: Decision support application: Based on the optimized model and data above, the TinyLSTM engine is trained and optimized using a credibility-enhanced dataset, enabling it to perform risk assessment, construction progress forecasting, resource optimization and design optimization during oil and gas pipeline construction based on real-time monitoring data, providing decision support for oil and gas pipeline construction.

[0154] S51: Define the training objectives and optimization problems:

[0155] Training objective: To enable the TinyLSTM engine to accurately perform risk assessment, construction progress prediction, resource optimization, and design optimization during oil and gas pipeline construction based on real-time monitoring data.

[0156] Optimization problem: Minimize the prediction error of the TinyLSTM engine while ensuring that its output conforms to the actual physical laws and engineering constraints of oil and gas pipeline construction;

[0157] S52: Training based on a credibility-enhanced dataset:

[0158] S521: Dataset preparation: Use the credibility-enhanced dataset X generated in step S326 enhanced As training data, this dataset contains processed pipeline monitoring data, construction-related data, etc.

[0159] S522: Model input definition: Define the input of the TinyLSTM engine as time series data X(t) = [x1(t), x2(t), ..., x n (t)] T , where x i (t) represents the value of the i-th monitoring parameter at time t;

[0160] S523: Model output definition: Risk assessment output Y risk (t), represents the probability of a risk event occurring at time t; the construction progress forecast output is Y progress (t), represents the construction progress at time t; resource optimization and allocation output Y resources (t), represents the demand for resources at time t; the design optimization output Y design (t), represents the optimization suggestion for the design scheme at time t;

[0161] S53: Model training and optimization algorithm:

[0162] S531: Loss function definition:

[0163] Risk assessment loss function: Where: N is the number of training samples, y risk,i is the actual risk label, is the risk probability predicted by the TinyLSTM engine;

[0164] Construction progress prediction loss function: Where: y progress,i The actual construction progress. is the construction progress predicted by the TinyLSTM engine;

[0165] Resource optimization and allocation loss function: Where: y resources,i is the actual resource demand, is the resource demand predicted by the TinyLSTM engine;

[0166] Design optimization loss function: Where: y design,i is the actual design optimization parameter, are the design optimization parameters predicted by the TinyLSTM engine;

[0167] S532: Overall loss function: L total =αL risk +βL progress +γL resources +δL design , where α, β, γ and δ are weight coefficients used to balance the losses of different tasks;

[0168] S533: Optimization algorithm: Use the Adam optimization algorithm to minimize the overall loss function L total , the parameters of the Adam optimization algorithm include the learning rate η, the first-order moment estimation parameter β1 and the second-order moment estimation parameter β2;

[0169] S534: Training process: Initialize the parameters of the TinyLSTM engine. For each training batch: Input the input data X(t) into the TinyLSTM engine and get the output Y risk (t),Y progress (t),Y resources (t) and Y design (t), calculate the overall loss function L total , use the back-propagation algorithm to calculate the gradient, and use the Adam optimization algorithm to update the model parameters, repeating the above steps until the model converges or reaches the preset number of training rounds;

[0170] S535: Model Evaluation and Validation: Use the validation dataset to evaluate the trained TinyLSTM engine and calculate its performance indicators (such as accuracy and mean squared error) for risk assessment, construction schedule prediction, resource optimization and allocation, and design optimization tasks. Based on the evaluation results, adjust and optimize the model, such as adjusting hyperparameters and adding regularization, to improve the model's performance and generalization ability.

[0171] S54: Model deployment and decision support: The trained and optimized TinyLSTM engine is deployed on edge devices to enable it to run and process real-time monitoring data. The edge devices collect pipeline monitoring data in real time and input it into the TinyLSTM engine. The TinyLSTM engine performs risk assessment, construction progress forecasting, resource optimization and allocation, and design optimization based on the real-time monitoring data, and outputs corresponding prediction results and decision recommendations. The prediction results and decision recommendations are provided to managers and decision makers of oil and gas pipeline construction, providing them with scientific decision support, helping them make more accurate and timely decisions, and improving the efficiency, safety, and reliability of oil and gas pipeline construction.

[0172] Through the above process, based on the optimized model and data, the TinyLSTM engine is trained and optimized using the credibility-enhanced dataset, enabling it to effectively provide decision support for oil and gas pipeline construction.

[0173] From the above description, it can be seen that the above-mentioned embodiments of the present invention achieve the following technical effects: data fusion and dynamic correction: introducing the spatiotemporal graph neural network (STGNN), fusing the pipeline strain sensor time series data and InSAR satellite remote sensing surface deformation data, constructing a dynamic offset compensation model, and correcting the coordinate drift caused by geological subsidence and pipeline displacement in real time, which significantly improves the accuracy and reliability of the data and provides a solid data foundation for subsequent analysis and decision-making.

[0174] Data credibility assessment and completion: Build a dynamic data credibility assessment chain, cross-validate multi-node sensor data based on a federated learning framework, generate confidence weights, and develop a generative adversarial network (GAN) completion module to generate synthetic data that conforms to physical laws for missing data in extreme working conditions, reducing the false alarm rate and enhancing the system's monitoring and fault diagnosis capabilities in complex environments.

[0175] Efficient model deployment and update: A layered distillation compression algorithm was designed to compress the physical information neural network (PINN) model trained in the cloud into a TinyLSTM engine suitable for edge device deployment. A model incremental update protocol was also developed to achieve efficient collaboration between the cloud and the edge. This reduces data transmission bandwidth usage, improves the system's real-time performance and response speed, and meets the real-time monitoring and intelligent analysis needs of oil and gas pipeline construction sites.

[0176] Intelligent Decision Support: Based on optimized models and credibility-enhanced datasets, the TinyLSTM engine is used to conduct risk assessment, construction progress forecasting, resource optimization and allocation, and design optimization during oil and gas pipeline construction. This provides scientific decision support for oil and gas pipeline construction, helps reduce construction costs, improve construction safety and reliability, and has significant economic and social benefits.

[0177] Those skilled in the art should understand that the discussion of any of the above embodiments is merely illustrative and is not intended to imply that the scope of the present invention is limited to these examples. Within the scope of the present invention, the technical features in the above embodiments or different embodiments may be combined, the steps may be implemented in any order, and there are many other variations of the different aspects of the present invention as described above, which are not provided in detail for the sake of simplicity.

[0178] The present invention is intended to cover all such substitutions, modifications and variations that fall within the broad scope of the specification. Therefore, any omissions, modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A digital twin-driven multi-source heterogeneous data fusion and intelligent analysis method for oil and gas pipeline construction, characterized by: The specific steps include: S1: Multi-source heterogeneous data acquisition: Collect multi-source heterogeneous data related to oil and gas pipeline construction, including pipeline strain sensor time series data, InSAR satellite surface deformation data, GIS geographic information, GPS positioning data and surrounding environment data; S2: Data fusion and dynamic compensation: Pipeline strain data and InSAR surface deformation data are fused through a spatiotemporal graph neural network to learn spatiotemporal correlation features and build a dynamic offset compensation model. This model is used to correct coordinate drift in real time and output corrected 3D coordinates and strain data. S3: Data credibility assessment and completion: Distributed nodes cross-validate sensor data to generate dynamic confidence weights. For missing data in extreme working conditions, physical models and historical data are used to generate synthetic data, and confidence weights are combined to generate a credible data set. S4: Edge Intelligent Optimization: A layered distillation algorithm is used to compress the cloud-based PINN model into a lightweight TinyLSTM engine, adapted for real-time computing on edge devices. Local training on edge devices generates incremental updates, which are aggregated in the cloud and then distributed to reduce bandwidth usage and achieve continuous model optimization. S5: Decision support application: Based on the optimized model and data described above, the TinyLSTM engine is trained and optimized using a credibility-enhanced dataset, enabling it to perform risk assessment, construction progress forecasting, resource optimization, and design optimization during oil and gas pipeline construction based on real-time monitoring data.

2. The digital twin-driven multi-source heterogeneous data fusion and intelligent analysis method for oil and gas pipeline construction according to claim 1 is characterized in that: The S2 specifically includes: S21: Build a spatiotemporal graph neural network model, taking the collected pipeline strain sensor time series data and InSAR satellite remote sensing surface deformation data as input. The spatiotemporal graph neural network model fuses the two data sources, learns the feature associations of the data in the spatiotemporal dimension, and builds a dynamic offset compensation model. S22: Using the dynamic offset compensation model, the coordinate data of the pipeline strain sensor is corrected in real time to compensate for the coordinate drift caused by geological settlement and pipeline displacement, and the corrected pipeline three-dimensional coordinate information and strain data are output.

3. The digital twin-driven multi-source heterogeneous data fusion and intelligent analysis method for oil and gas pipeline construction according to claim 2 is characterized in that: The S21 specifically includes: S211: Data preprocessing and fusion: The collected pipeline strain sensor time series data is S(t), and the InSAR satellite remote sensing surface deformation data is I(t). The two data are aligned in the time dimension, and the missing time points are filled using the interpolation method. The aligned pipeline strain sensor time series data and the InSAR satellite remote sensing surface deformation data are fused to construct a fused data matrix X(t) = [S(t); I(t)]; S212: Spatiotemporal graph construction: Define the graph structure G = (V, E), where V represents the set of nodes and E represents the set of edges between nodes; calculate the distance matrix D between nodes, determine the connection relationship between nodes based on the distance threshold, and if the distance between two nodes is less than the threshold, establish an edge between them; construct the adjacency matrix A of the graph, where the element A ij represents the connection weight between node i and node j; S213: Spatiotemporal graph convolution operation: In spatiotemporal graph neural networks, graph convolution operations are used to capture spatial feature correlations between nodes; for capturing feature correlations in the time dimension, one-dimensional convolution operations are used; S214: Dynamic offset compensation model construction: In the output layer of STGNN, the time attenuation coefficient α and the spatial offset coefficient β are introduced to construct a dynamic offset compensation model.

4. The digital twin-driven multi-source heterogeneous data fusion and intelligent analysis method for oil and gas pipeline construction according to claim 3 is characterized in that: The S22 specifically includes: S221: Define coordinate data and strain data: Assume that the original three-dimensional coordinate data of the pipeline strain sensor is P(t), and the original strain data is Q(t) = {ε axial (t),ε hoop (t)}, where ε axial (t) represents the axial strain, ε hoop (t) represents hoop strain; S222: Obtaining an offset compensation amount: Obtaining an offset compensation amount ΔP(t) according to the dynamic offset compensation model in step S214; S223: Coordinate data correction: Corrected pipeline three-dimensional coordinate information P corrected (t) = P(t) + ΔP(t); S224: Strain data correction: Introduce strain correction coefficient k ε , the corrected strain data S225: Model training and parameter update: Define the loss function L for training the dynamic offset compensation model, combining the coordinate correction error and the strain correction error; use the optimization algorithm to update the model parameters α and β to minimize the loss function L.

5. The digital twin-driven multi-source heterogeneous data fusion and intelligent analysis method for oil and gas pipeline construction according to claim 1 is characterized in that: The S3 specifically includes: S31: Build a dynamic data credibility evaluation chain. Based on the federated learning framework, cross-validate sensor data distributed across multiple nodes and generate confidence weights for each sensor data. S32: Develop a generative adversarial network completion module. For missing pipeline data under extreme working conditions, use the existing pipeline physical model and historical data to generate synthetic data that conforms to physical laws through the GAN module to complete the missing data. The completed data is weighted based on the confidence weight to generate a dataset with enhanced credibility.

6. The digital twin-driven multi-source heterogeneous data fusion and intelligent analysis method for oil and gas pipeline construction according to claim 5 is characterized in that: The S31 specifically includes: S311: Define sensor data: Assume that the sensor data distributed on multiple nodes is D = {D1, D2, ..., D m }, where m represents the number of sensor nodes, D i ={d i1 ,d i2 ,...,d in } represents the data collected by the i-th node at n time steps; S312: Federated learning framework initialization: In the federated learning framework, set up a central server and multiple clients, define the global model parameter θ, initialize it to θ0, set the learning rate α, momentum coefficient β and model decay coefficient γ; S313: Data consistency evaluation: For each sensor node i, calculate the consistency score C of its data with the data of other nodes i ; S314: Data stability evaluation: Calculate the stability score S of each sensor node i data i , based on the time series variation characteristics of data; S315: Confidence weight calculation: Combine the consistency and stability scores to generate the confidence weight W of each sensor data i , fused through quadratic functions: S316: Federated learning model update: Each client uses local data D i and the current global model parameters θ, calculate the gradient And use the gradient descent method with momentum to update the local model parameters; the client will update the model parameters θ i ′ is sent to the central server, which aggregates the parameters of all clients and updates the global model parameters.

7. The digital twin-driven multi-source heterogeneous data fusion and intelligent analysis method for oil and gas pipeline construction according to claim 6 is characterized in that: The S32 specifically includes: S321: Define the GAN module structure: The GAN module consists of a generator G and a discriminator D. The input of the generator G is a random noise vector z and some known pipeline data X. known , the output is the completed pipeline data X generated , the input of the discriminator D is the real pipeline data X real Or the data X generated by the generator generated , the output is the judgment result P real , represents the probability that the input data is true data; S322: Combining pipeline physical model and historical data: Define pipeline physical model M, which is based on physical laws and combines pipeline data with physical constraints; collect historical data H, including pipeline data collected in the past under similar operating conditions, for training the GAN module; S323: Generator training: The goal of the generator G is to minimize the difference between the generated data and the real data while satisfying the physical model constraints. Its loss function L G ; S324: Discriminator training: The goal of the discriminator D is to distinguish between real data and generated data, and its loss function L D ; S325: Generate synthetic data: After training is completed, for the missing pipeline data X under extreme working conditions missing , use the generator G to generate the complete dataset X complete ; S326: Combined with confidence weight weighting processing: For the completed data X complete , combined with the confidence weight W obtained in step S315 i , generate a credibility-enhanced dataset X enhanced .

8. The digital twin-driven multi-source heterogeneous data fusion and intelligent analysis method for oil and gas pipeline construction according to claim 1 is characterized in that: The S4 specifically includes: S41: Design a hierarchical distillation compression algorithm to compress the physical information neural network model trained in the cloud into a TinyLSTM engine suitable for edge device deployment. This allows the edge device to run the TinyLSTM engine in real time to perform pipeline deformation and buckling calculations. S42: Develop a model incremental update protocol. The edge node uses the newly collected local data to locally train the TinyLSTM engine, and sends the incremental model update information obtained from the local training to the cloud. The cloud aggregates and iterates the incremental update information of multiple edge nodes to generate a new model update package, and then sends the update package to the edge node to realize the incremental update of the model on the edge device and reduce the data transmission bandwidth usage.

9. The digital twin-driven multi-source heterogeneous data fusion and intelligent analysis method for oil and gas pipeline construction according to claim 8 is characterized in that: The S41 specifically includes: S411: Define the structure of the layered distillation compression algorithm: Cloud-based physical information neural network model: The model structure includes multiple fully connected layers, activation function layers, and physical constraint layers; Edge TinyLSTM engine: The TinyLSTM engine takes as input the time series of pipeline physical parameters and outputs real-time pipeline deformation and buckling prediction values. S412: Layered Distillation Compression Algorithm: Knowledge distillation loss function: Define the knowledge distillation loss function L distill , used to measure the difference between the output of the PINN model and the TinyLSTM engine; Hierarchical distillation process: the output Y of the PINN model PINN As the output of the teacher model, the output Y of the TinyLSTM engine is TinyLSTM As the output of the student model, during the training process, the knowledge distillation loss function L is minimized. distill , so that the TinyLSTM engine can learn the knowledge of the PINN model; Model compression steps: Prune the PINN model weights to remove unimportant weights and reduce the number of model parameters; quantize the PINN model weights from floating point numbers to low-precision integers to reduce model storage space; use the knowledge distillation method to train the TinyLSTM engine using the output of the pruned and quantized PINN model as a guide; S413: Edge Device Deployment: TinyLSTM engine optimization: Further optimize the trained TinyLSTM engine, such as model pruning and quantization. Real-time deformation and buckling calculation: The optimized TinyLSTM engine is deployed on edge devices. The edge devices collect real-time pipeline physical parameters and input them into the TinyLSTM engine for real-time pipeline deformation and buckling calculation.

10. The digital twin-driven multi-source heterogeneous data fusion and intelligent analysis method for oil and gas pipeline construction according to claim 9 is characterized in that: The S42 specifically includes: S421: Local training of edge nodes: defining local training data, updating local models, calculating local gradients, and generating incremental update information; S422: Cloud aggregation iteration: collect incremental update information, define aggregation weights, aggregate incremental update information, and update global model parameters; S423: Model update distribution: The cloud distributes the updated global model parameters Packaged into model update package P update , the cloud will P update Send it to each edge node. After receiving the update package, the edge node replaces the local model parameters with Complete the incremental update of the model.

Images