Multi-dimensional information interconnection and autonomous evolution collaborative pipeline full-state safety assessment method

Abstract

The present application belongs to the technical field of pipeline safety assessment, and discloses a pipeline full-state safety assessment method based on multi-dimensional information interconnection and autonomous evolution coordination. The high-order relationship of data is captured through dual hypergraph reasoning of instance-level hypergraph and modal-level hypergraph, and efficient interconnection is realized. The modal-level and instance-level hypergraph information features are extracted through hypergraph information propagation, the high-order correlation is mined through dual graph information aggregation, and the cross-modal and cross-instance consistent information and exclusive information are output after feature reorganization, which promotes the deep fusion of multi-dimensional data and provides high-quality data support for subsequent pipeline full-state safety assessment. The two-level autonomous evolution mechanism of intra-class gradual calibration and inter-class knowledge transfer is respectively adapted to the scenes of slight fluctuation and significant change of deep sea environment: within the domain, the dual-branch feature extraction and dynamic weight adjustment are realized to accurately adapt to the environmental perturbation; between the domains, the spatiotemporal feature clustering and cross-domain knowledge transfer are completed to realize the dynamic optimization of model parameters without manual intervention, and the dynamic environment self-adaptation is realized.

Description

Technical Field

[0001] This invention relates to the field of pipeline safety assessment technology, and in particular to a pipeline full-state safety assessment method that combines multi-dimensional information interconnection and autonomous evolution. Background Technology

[0002] As the core transportation hub for offshore oil and gas resource development, deep-sea pipelines are the "energy arteries" connecting deep-sea oil and gas fields with land terminals. Their safe and stable operation is directly related to the sustainability of the marine ecological environment.

[0003] The deep-sea environment is characterized by extreme conditions such as high pressure, low temperature, strong corrosion, and frequent geological activity. Pipelines in this environment face multiple risks, including corrosion fatigue, structural damage, and third-party sabotage. Leaks not only result in massive waste of oil and gas resources and economic losses, but also disrupt the marine ecological balance, threatening the survival of marine life and the livelihoods of coastal residents. Furthermore, deep-sea pipelines are difficult and costly to repair, and the aftermath of accidents is lengthy, potentially causing long-term impacts on the stability of regional energy supply.

[0004] Currently, there are three main methods for assessing pipeline safety status: methods based on deep learning models built from data, methods based on mechanistic models, and methods based on digital twins.

[0005] The main steps of the data-based deep learning model building method are: (1) data preprocessing; (2) feature extraction and feature association of data; (3) artificial intelligence model evaluation; (4) reliable decision output. The data-based deep learning model building method has certain limitations: (1) when associating features of multidimensional data, it is difficult to capture the complex high-order relationships of different modal information; (2) in the face of dynamic and changing environments, deep learning models often have insufficient generalization ability, especially when the environment changes greatly, the model is not applicable because it cannot evolve autonomously; (3) in extreme environments, environmental noise, signal attenuation and other factors lead to large uncertainty of existing state evaluation and insufficient reliability of results.

[0006] The main steps of the mechanism-based model method are: (1) analyzing the physical mechanism of the detection method; (2) establishing a mathematical model; (3) optimizing the model parameters; and (4) comparing the results of the mechanism model with the actual situation to detect pipeline defects. The mechanism-based model method has certain limitations: (1) modeling depends on the understanding of physical mechanisms, and complex phenomena that are not clearly defined are difficult to describe accurately; (2) the generalization ability is weak, and it is necessary to remodel to adapt to complex working conditions (such as deep sea and multi-media corrosion); (3) simplifying assumptions can easily ignore the coupling effect of multiple factors, resulting in limited model accuracy.

[0007] The main steps of the digital twin-based method are: (1) constructing a virtual-real mapping twin of the pipeline; (2) fusing multi-source data; (3) dynamically simulating and deducing the pipeline status; and (4) assessing risks and predicting maintenance needs. The digital twin-based method also has certain limitations: (1) when real-time requirements are high, data transmission delay and large simulation calculation volume may lead to lag in dynamic response; (2) the initial construction and operation and maintenance costs are high, and the threshold for sensor deployment and system update and maintenance is high; (3) in extreme environments, the robustness of the virtual-real mapping is easily affected, and the model adaptability is insufficient. Summary of the Invention

[0008] To address the issues of information silos, delayed dynamic responses, inability of models to evolve autonomously, and insufficient reliability of assessments caused by the extreme, complex, and uncertain nature of the deep-sea environment, this invention proposes a pipeline full-state safety assessment method that integrates multi-dimensional information interconnection and autonomous evolution. This method enables accurate pipeline status assessment and ensures the safe and reliable operation of deep-sea pipelines.

[0009] The technical solution of this invention is as follows: A pipeline full-state safety assessment method based on multi-dimensional information interconnection and autonomous evolution collaboration, comprising the following steps:

[0010] Step 1: Acquire three types of sensor data: magnetic flux leakage, ultrasonic, and electromagnetic eddy current, perform preprocessing, and obtain preprocessed multidimensional data;

[0011] Step 2: Construct a multi-dimensional sensor information interconnection module. Use the hypergraph modeling method to interconnect the preprocessed multi-dimensional data to obtain associated features.

[0012] Step 3: Build a multi-scene feature mining module to perform multi-scene feature mining based on the current environment state representation vector and related features;

[0013] Step 4: Construct an evaluation model based on the multi-dimensional sensor information interconnection module and the multi-scenario feature mining module; collect standard environmental data and target environmental data to train the evaluation model and obtain the pipeline full-state safety evaluation model.

[0014] The preprocessing includes data dimensionality reduction and purification using ensemble empirical mode decomposition and Z-score standardization.

[0015] After standardization and outlier removal for each sensor data format, ensemble empirical mode decomposition is performed. Each sensor data is decomposed into five IMF components and one residual component according to frequency. For magnetic flux leakage sensing data, the data at the defect location corresponding to the fourth IMF component is taken, while for ultrasonic sensing data and electromagnetic eddy current sensing data, the data at the defect location corresponding to the fifth IMF component is taken respectively. Z-score standardization is performed on the IMF components corresponding to the defect location data to obtain preprocessed multidimensional data.

[0016] The hypergraph modeling method uses either a single instance or modal information as nodes; describes the complex relationships between nodes through hyperedges, and establishes instance-level hypergraphs and modal-level hypergraphs respectively; and mines the association features of instance-level hypergraphs and modal-level hypergraphs by using hypergraph information propagation and dual-graph information aggregation.

[0017] The example is a sample of a pipe defect, including its length, width, depth, and relative location; this is known data.

[0018] The modal information is preprocessed multidimensional data.

[0019] The instance-level hypergraph establishes cross-modal instance associations using each pipeline defect sample as a node;

[0020] The modal-level hypergraph uses preprocessed multidimensional data as nodes to characterize semantic associations and responses to environmental changes throughout the entire data acquisition process, and describes modal associations across instances.

[0021] The instance-level hypergraph and the modal-level hypergraph perform hypergraph information propagation respectively, with the instance-level hypergraph's hypergraph information propagation being dominant;

[0022] Propagation loss of hypergraph information propagation for: ;

[0023] in, For instance-level hypergraphs, ideal features are based on association labels. Ideal features for modal-level hypergraphs based on associated labels; For instance-level hypergraphs in the first... T Propagation characteristics of information after propagation in a hypergraph; For modal-level hypergraphs in the first... T Propagation characteristics of information after propagation in a hypergraph; MSE Mean square error;

[0024] The core of hypergraph information propagation in instance-level hypergraphs is feature-weighted fusion based on hyperedge weights, using the following formula:

[0025]

[0026] in, For instance-level hypergraphs t The propagation characteristics of information after propagation in a hypergraph. t When it is 0, This provides detailed information about pipeline defects after Z-score standardization. The degree matrix of the instance-level hypergraph; Let be the adjacency matrix of the instance-level hypergraph, representing the connection relationships between nodes in the instance-level hypergraph; For instance-level hypergraphs tThe weight matrix of the hyperedge. For bias terms; The propagation coefficients of the instance-level hypergraph; For activation functions;

[0027] After each round of hypergraph information propagation, the weights of hyperedges in the instance-level hypergraph are adjusted based on node feature similarity, and the update formula is used. , The cosine similarity function;

[0028] The hypergraph information propagation of modal-level hypergraphs uses the following formula:

[0029] ;

[0030] For modal-level hypergraphs t Propagation characteristics of information after propagation in a hypergraph; The propagation coefficients of the modal-level hypergraph; Let be the degree matrix of the modal-level hypergraph; Let be the adjacency matrix of the modal hypergraph, representing the connection relationships between nodes in the modal hypergraph; For modal-level hypergraphs t Wheel hyperedge weight matrix; For bias terms;

[0031] After each round of hypergraph information propagation, the hyperedge weights in the modal hypergraph are dynamically adjusted based on environmental parameters, using the following formula: ; The environmental impact score is calculated based on ocean current velocity and pressure, with a value range of [0, 0.3]. This is the adjustment coefficient.

[0032] The dual-graph information aggregation uses an attention mechanism and feature recombination to fuse the propagation features of instance-level hypergraphs and modality-level hypergraphs, extracting consistency features and specific features.

[0033] The attention mechanism is a dual-graph attention layer. It calculates the association weights between instance-level hypergraph nodes and modal-level hypergraph nodes, and calculates the cross-graph fusion features of the instance-level and modal-level hypergraphs based on the association weights. The propagation features of the instance-level hypergraph and the cross-graph fusion features are then recombined, and consistent features are output through two branches. and exclusive features These together describe the specific relationship between sensor information and defect information, and are collectively referred to as associated features;

[0034] The consistency feature The acquisition process is as follows:

[0035] ;

[0036] in,[ m ] represents the m-th instance-level hypergraph node. Indicates feature splicing; This represents a one-dimensional convolutional layer; Indicates cross-graph fusion features;

[0037] The specific features The acquisition process is as follows:

[0038]

[0039] in, It is a mean function;

[0040] The loss function for hypergraph association optimization consists of three parts: propagation loss, etc. fusion loss Mission losses , For classifiers, For real labels, Calculate the cross-entropy;

[0041] During joint training, the propagation loss coefficient, fusion loss coefficient, and task loss coefficient were set to 0.4, 0.3, and 0.3, respectively.

[0042] The multi-scene feature mining module includes intra-class progressive calibration and inter-class knowledge transfer;

[0043] The intra-class progressive calibration specifically involves obtaining the current environment state representation vector, extracting the relationship between the current environment state representation vector and associated features using a dual-branch feature extraction structure, and constructing feature consistency constraints based on the extracted relationship.

[0044] Multi-source environmental parameters, including temperature, are acquired through sensors around the pipeline. T ,pressure P and flow rate V After standardizing the multi-source environmental parameters, a feature fusion algorithm is used to generate a current environmental state representation vector that highly matches the associated features. E The formula is as follows:

[0045]

[0046] in, , , The weights of environmental parameters determined based on correlation analysis of pipeline safety assessment, Norm ( ) is a standardized function;

[0047] The dual-branch feature extraction structure includes a global view branch and a local view branch; the global view branch provides consistent features. and exclusive features After positional encoding, the data is processed by a transformer encoder followed by global average pooling to obtain long-range dynamic features, which are then used as the global feature vector F. g Local view branches have consistent features. and exclusive features After performing convolution operations using a 3×1 small convolution kernel, local convolution features are obtained. The local convolution features are then compared with the environmental state representation vector. E The cosine similarity correlation was used, and a gating mechanism was employed to filter local features strongly correlated with environmental perturbations, which were then used as the local feature vector F. l ; the global feature vector F g With local eigenvector F l The features are concatenated to generate a fused feature vector F=[F g F l ];

[0048] To align the fused feature vector with the semantic reference features of the samples obtained through pre-training, a semantically aware constrained loss function is designed. To apply feature consistency constraints, the formula is as follows:

[0049]

[0050] in, N For the sample size, cos( ) is the cosine similarity function. Let be the fused feature vector of the i-th pipeline defect sample. The sample semantic reference features, which are semantically consistent with the i-th pipeline defect sample, are obtained by pre-training on standard environmental data under a stable and invariant environment.

[0051] The inter-class knowledge transfer includes spatiotemporal feature clustering and hyperdimensional space constraints;

[0052] The spatiotemporal feature clustering is as follows;

[0053] First, spatiotemporal alignment preprocessing is performed on consistency features and specific features to extract temporal and spatial dimension features, and a spatiotemporal feature vector X=[X t X s ], X t X is a subset of time features. s A subset of spatial features;

[0054] The K-means clustering algorithm is adopted, which measures feature similarity based on cosine similarity and determines the number of clusters K according to the elbow rule to generate multi-class feature representation primitives.

[0055] The feature representation primitives and the corresponding sample semantic reference features are mapped to a hyperdimensional space, and constrained optimization is performed in the hyperdimensional space.

[0056] Intragroup compactness constraints Reduce the distance between similar feature representation primitives, C k Let k be the set of feature representation primitives. n k The number of samples in the set. As the class center; constrained by between-group separability Expand the distributional differences of heterogeneous feature representation primitives. The global feature representation primitive center is established; spatiotemporal correlation constraints are introduced. Corr is the environmental evolution rate factor. () represents the correlation coefficient function;

[0057] Based on the triple constraints of intra-group compactness, inter-group separability, and spatiotemporal correlation, the knowledge of the distribution law of the feature representation primitives of the source class is transferred to the hyperdimensional space of the target domain.

[0058] The pipeline full-state safety assessment model is obtained by training through a total loss function:

[0059]

[0060] in 、 、 To constrain the weights, the initial values ​​are all set to 1;

[0061] First, pre-training is performed using labeled standard environmental data. During pre-training, the multi-scenario feature mining module is turned off to obtain a pre-trained pipeline full-state safety assessment model.

[0062] Freeze all parameters in the pre-trained evaluation model, unlock the multi-scene feature mining module, and train using target environment data;

[0063] Using target environment data as training samples, based on the total loss function L total The parameters of the multi-scene feature mining module are updated iteratively using the mini-batch gradient descent algorithm. Each iteration only uses incremental data fragments of the target environment data and does not retrain the entire pre-training pipeline full-state safety assessment model.

[0064] Set the parameter update threshold to a decrease in the loss function of less than 10. -4 The current update is stopped when the pipeline's full-state safety assessment model parameters are dynamically optimized.

[0065] The beneficial effects of this invention are as follows: This invention captures high-order relationships in data through dual hypergraph inference at the instance level and modality level, achieving efficient interconnection. The instance-level hypergraph uses pipeline defect samples as nodes to capture fine-grained cross-modal correlations; the modality-level hypergraph uses sensor modal information as nodes to characterize global semantic correlations and environmental responses. Then, the hypergraph correlations are analyzed, and modality-level and instance-level hypergraph information features are extracted through hypergraph information propagation. High-order correlations are mined through dual-graph information aggregation, and after feature recombination, cross-modal and cross-instance consistent information and specific information are output, promoting deep fusion of multi-dimensional data and providing high-quality data support for subsequent pipeline full-state safety assessments.

[0066] This invention proposes a two-stage autonomous evolution mechanism of "intra-class progressive calibration and inter-class knowledge transfer," which differs from existing model transfer methods based on steady-state switching. It can adapt to scenarios with slight fluctuations and significant changes in the deep-sea environment: within the domain, it achieves accurate adaptation to environmental perturbations through bi-branch feature extraction and dynamic weight adjustment; between the domains, it completes dynamic optimization of model parameters through spatiotemporal feature clustering and cross-domain knowledge transfer, achieving dynamic environmental adaptation without human intervention. Attached Figure Description

[0067] Figure 1 A flowchart of a pipeline full-state safety assessment method based on multi-dimensional information interconnection and autonomous evolution collaboration.

[0068] Figure 2 This is a flowchart of a multi-dimensional sensor information interconnection module. Detailed Implementation

[0069] This invention mainly uses model evaluation methods based on three types of sensor data—magnetic leakage, ultrasonic, and electromagnetic eddy current—to conduct a full-state safety assessment of pipelines.

[0070] This invention presents a pipeline full-state safety assessment method based on multi-dimensional information interconnection and autonomous evolution collaboration, primarily targeting three types of sensor data: magnetic leakage flux, ultrasonic, and electromagnetic eddy current. Unlike existing single-dimensional data analysis methods, this invention proposes a dimensionality reduction and purification mechanism for multi-dimensional sensor data under strong interference and an information interconnection method. It explores the interconnection characteristics of multi-dimensional sensor data, promotes the deep integration and collaboration of multi-dimensional sensor data, and thus provides data support for the construction of a pipeline full-state safety assessment system.

[0071] A pipeline full-state safety assessment method based on multi-dimensional information interconnection and autonomous evolution collaboration includes the following steps:

[0072] Step 1: Acquire three types of sensor data: magnetic flux leakage, ultrasonic, and electromagnetic eddy current, perform preprocessing, and obtain multidimensional data after multidimensional preprocessing.

[0073] Step 2: To accurately capture the complex high-order relationships between different data types, thereby overcoming information bottlenecks and achieving efficient information integration and sharing, a hypergraph modeling method is used to interconnect preprocessed multidimensional data with multidimensional sensor information to obtain associated features.

[0074] Step 3: Perform multi-scene feature mining based on the current environment state representation vector and associated features;

[0075] Step 4: Construct an evaluation model based on the multi-dimensional sensor information interconnection module and the multi-scenario feature mining module; collect standard environmental data and target environmental data to train the evaluation model and obtain the pipeline full-state safety evaluation model.

[0076] The multidimensional data preprocessing includes dimensionality reduction and purification of data using ensemble empirical mode decomposition and Z-score standardization;

[0077] After standardization and outlier removal for each sensor data format, ensemble empirical mode decomposition is performed. Each sensor data is decomposed into five IMF components and one residual component according to frequency. For magnetic flux leakage sensing data, the data at the defect location corresponding to the fourth IMF component is taken, while for ultrasonic sensing data and electromagnetic eddy current sensing data, the data at the defect location corresponding to the fifth IMF component is taken respectively. Z-score standardization is performed on the IMF components corresponding to the defect location data to obtain preprocessed multidimensional data.

[0078] The hypergraph modeling method uses either a single instance or modal information as vertices; describes the complex relationships between vertices through hyperedges, and establishes instance-level and modal-level hypergraphs respectively; and mines the association features of instance-level and modal-level hypergraphs by using hypergraph information propagation and dual-graph information aggregation.

[0079] The example is a sample of a pipe defect, including its length, width, depth, and relative location, which are known data;

[0080] The modal information is preprocessed multidimensional data.

[0081] The instance-level hypergraph establishes cross-modal instance associations using each pipeline defect sample as a node;

[0082] The modal-level hypergraph uses preprocessed multidimensional data as nodes to characterize semantic associations and responses to environmental changes throughout the entire data acquisition process, and describes modal associations across instances.

[0083] The instance-level hypergraph and the modal-level hypergraph perform hypergraph information propagation respectively, with the instance-level hypergraph's hypergraph information propagation being dominant;

[0084] Loss function for hypergraph information propagation for: ;

[0085] in, For instance-level hypergraphs, ideal features are based on association labels. Ideal features for modal-level hypergraphs based on associated labels; For instance-level hypergraphs in the first... T The propagation characteristics of wheels, For modal-level hypergraphs in the first... T The propagation characteristics of wheels; MSE Mean square error;

[0086] The core of hypergraph information propagation in instance-level hypergraphs is feature-weighted fusion based on hyperedge weights, using the following formula:

[0087]

[0088] in, For instance-level hypergraphs t The propagation characteristics of information after propagation in a hypergraph. t When it is 0, This provides detailed information about pipeline defects after Z-score standardization. The degree matrix of the instance-level hypergraph; Let be the adjacency matrix of the instance-level hypergraph, representing the connection relationships between nodes in the instance-level hypergraph; For instance-level hypergraphs t The weight matrix of the hyperedge. For bias terms; The propagation coefficients of the instance-level hypergraph; For activation functions;

[0089] After each round of hypergraph information propagation, the weights of hyperedges in the instance-level hypergraph are adjusted based on node feature similarity, and the update formula is used. , The cosine similarity function;

[0090] The hypergraph information propagation of modal-level hypergraphs uses the following formula:

[0091] ;

[0092] For modal-level hypergraphs t Propagation characteristics of information after propagation in a hypergraph; The propagation coefficients of the modal-level hypergraph; The degree matrix of the modal-level hypergraph; Let be the adjacency matrix of the modal hypergraph, representing the connection relationships between nodes in the modal hypergraph; For modal-level hypergraphs t Wheel hyperedge weight matrix; For bias terms;

[0093] After each round of hypergraph information propagation, the hyperedge weights in the modal hypergraph are dynamically adjusted based on environmental parameters, using the following formula: ; The environmental impact score is calculated based on ocean current velocity and pressure, with a value range of [0, 0.3]. This is the adjustment coefficient.

[0094] The dual-graph information aggregation, through attention mechanism and feature recombination, fuses the propagation features of instance-level hypergraphs and modal-level hypergraphs, extracts consistency features and specific features, and jointly describes the complex high-order relationship between instance-level hypergraphs and modal-level hypergraphs.

[0095] The attention mechanism is a dual-graph attention layer. It calculates the association weights between instance-level hypergraph nodes and modal-level hypergraph nodes, and calculates the cross-graph fusion features of instance-level hypergraph nodes based on the association weights. The propagation features of instance-level hypergraph nodes and cross-graph fusion features are then recombined, and consistent features are output through two branches. and exclusive features These together describe the specific relationship between sensor information and defect information, and are collectively referred to as associated features;

[0096] The consistency feature The acquisition process is as follows:

[0097] ;

[0098] Where [m] represents the m-th instance-level hypergraph node, Indicates feature splicing; This represents a one-dimensional convolutional layer; Indicates cross-graph fusion features;

[0099] The specific features The acquisition process is as follows:

[0100]

[0101] in, It is a mean function;

[0102] The loss function for hypergraph association optimization consists of three parts: propagation loss, etc. fusion loss Mission losses , For classifiers, For real labels, Calculate the cross-entropy;

[0103] During joint training, the propagation loss coefficient, fusion loss coefficient, and task loss coefficient were set to 0.4, 0.3, and 0.3, respectively.

[0104] The multi-scenario feature mining includes intra-class progressive calibration and inter-class knowledge transfer;

[0105] The intra-class progressive calibration specifically involves acquiring the current environmental condition features, using a dual-branch feature extraction structure to extract the relationship between the current environmental condition features and related features, and constructing feature consistency constraints based on the extracted relationship.

[0106] Multi-source environmental parameters, including temperature (T), pressure (P), and flow velocity (V), are acquired through sensors around the pipeline. After standardizing these parameters, a feature fusion algorithm is used to generate a current environmental state representation vector that highly matches the associated features. E The formula is as follows:

[0107]

[0108] in, , , The weights of environmental parameters determined based on correlation analysis of pipeline safety assessment, Norm ( ) is a standardized function;

[0109] The dual-branch feature extraction structure includes a global view branch and a local view branch; the global view branch provides consistent features. and exclusive features After positional encoding, the data is processed by a transformer encoder followed by global average pooling to obtain long-range dynamic features, which are then used as the global feature vector F. g Local view branch pairs and After performing convolution operations using a 3×1 small convolution kernel, local convolution features are obtained. The local convolution features are then compared with the environmental state representation vector. E The cosine similarity correlation was used, and a gating mechanism was employed to filter local features strongly correlated with environmental perturbations, which were then used as the local feature vector F. l ; the global feature vector F g With local eigenvector F l The features are concatenated to generate a fused feature vector F=[F g F l ];

[0110] The inter-class knowledge transfer includes spatiotemporal feature clustering and hyperdimensional space constraints;

[0111] The spatiotemporal feature clustering is as follows;

[0112] First, spatiotemporal alignment preprocessing is performed on consistency features and specific features to extract temporal and spatial dimension features, and a spatiotemporal feature vector X=[X t Xs ], X t X is a subset of time features. s A subset of spatial features;

[0113] The K-means clustering algorithm is adopted, which measures feature similarity based on cosine similarity and determines the number of clusters K according to the elbow rule to generate multi-class feature representation primitives.

[0114] The feature representation primitives and the corresponding sample semantic reference features are mapped to a hyperdimensional space, and constrained optimization is performed in the hyperdimensional space.

[0115] Intragroup compactness constraints Reduce the distance between similar feature representation primitives, C k Let k be the set of feature representation primitives. n k The number of samples in the set. Using class centers and between-group separability constraints Expand the distributional differences of heterogeneous feature representation primitives. The global feature representation primitive center is established; spatiotemporal correlation constraints are introduced. Corr is the environmental evolution rate factor. () represents the correlation coefficient function;

[0116] Based on the triple constraints of intra-group compactness, inter-group separability, and spatiotemporal correlation, the knowledge of the distribution law of the feature representation primitives of the source class is transferred to the hyperdimensional space of the target domain.

[0117] The pipeline full-state safety assessment model is trained using a total loss function:

[0118]

[0119] in , , To constrain the weights, the initial values ​​are all set to 1;

[0120] First, pre-training is performed using labeled sensor data in a standard environment. During pre-training, the multi-scene feature mining module is turned off to obtain a pre-trained evaluation model.

[0121] Freeze all parameters in the pre-trained evaluation model, unlock the multi-scene feature mining module, and train using target environment data;

[0122] Using data from the target class as training samples, based on the total loss function L total The parameters of the multi-scene feature mining module are updated iteratively using the mini-batch gradient descent algorithm. Each iteration only uses incremental data fragments of the target class, without the need to retrain the entire evaluation model.

[0123] Set the parameter update threshold to a decrease in the loss function of less than 10. -4 The current update cycle is stopped when the time is right, allowing for dynamic optimization of the evaluation model parameters.

[0124] This invention proposes a pipeline full-state safety assessment method based on multi-dimensional information interconnection and autonomous evolution. By constructing instance-level and modal-level dual hypergraphs and designing hypergraph information propagation and dual-graph collaborative aggregation modules, this invention fully captures the rich structural information contained within the hypergraphs, achieving the mining and optimization of high-order correlations in multi-dimensional data. Addressing the issue of fixed weights in the assessment model, this invention adopts a strategy of slight fluctuation feature calibration and significant change knowledge transfer, thereby establishing a dynamic interconnection mechanism between the deep-sea environment and the assessment model, enabling real-time optimization of the assessment model for dynamic and unknown scenarios.

Claims

1. A pipeline full-state safety assessment method based on multi-dimensional information interconnection and autonomous evolution collaboration, characterized in that, The steps include the following: Step 1: Acquire three types of sensor data: magnetic flux leakage, ultrasonic, and electromagnetic eddy current, perform preprocessing, and obtain preprocessed multidimensional data; Step 2: Construct a multi-dimensional sensor information interconnection module. Use the hypergraph modeling method to interconnect the preprocessed multi-dimensional data to obtain associated features. Step 3: Build a multi-scene feature mining module to perform multi-scene feature mining based on the current environment state representation vector and related features; Step 4: Construct an evaluation model based on the multi-dimensional sensor information interconnection module and the multi-scenario feature mining module; collect standard environmental data and target environmental data to train the evaluation model and obtain the pipeline full-state safety evaluation model; The hypergraph modeling method uses either a single instance or modal information as nodes; describes the complex relationships between nodes through hyperedges, and establishes instance-level hypergraphs and modal-level hypergraphs respectively; and mines the association features of instance-level hypergraphs and modal-level hypergraphs by using hypergraph information propagation and dual-graph information aggregation. The example is a sample of a pipe defect, including its length, width, depth, and relative location; this data is known. The modal information is preprocessed multidimensional data; The instance-level hypergraph establishes cross-modal instance associations using each pipeline defect sample as a node; The modal-level hypergraph uses preprocessed multidimensional data as nodes to characterize semantic associations and responses to environmental changes throughout the entire data acquisition process, and describes modal associations across instances.

2. The pipeline full-state safety assessment method based on multi-dimensional information interconnection and autonomous evolution collaboration according to claim 1, characterized in that, The preprocessing includes data dimensionality reduction and purification using ensemble empirical mode decomposition and Z-score standardization. After standardization and outlier removal for each sensor data format, ensemble empirical mode decomposition is performed. Each sensor data is decomposed into five IMF components and one residual component according to frequency. For magnetic flux leakage sensing data, the data at the defect location corresponding to the fourth IMF component is taken, while for ultrasonic sensing data and electromagnetic eddy current sensing data, the data at the defect location corresponding to the fifth IMF component is taken respectively. Z-score standardization is performed on the IMF components corresponding to the defect location data to obtain preprocessed multidimensional data.

3. The pipeline full-state safety assessment method based on multi-dimensional information interconnection and autonomous evolution coordination according to claim 1, characterized in that, The instance-level hypergraph and the modal-level hypergraph perform hypergraph information propagation respectively, with the instance-level hypergraph's hypergraph information propagation being dominant; Propagation loss of hypergraph information propagation for: ; in, For instance-level hypergraphs, ideal features are based on association labels. Ideal features for modal-level hypergraphs based on associated labels; For instance-level hypergraphs in the first... T Propagation characteristics of information after propagation in a hypergraph; For modal-level hypergraphs in the first... T Propagation characteristics of information after propagation in a hypergraph; MSE Mean square error; The core of hypergraph information propagation in instance-level hypergraphs is feature-weighted fusion based on hyperedge weights, using the following formula: in, For instance-level hypergraphs t The propagation characteristics of information after propagation in a hypergraph. t When it is 0, This provides detailed information about pipeline defects after Z-score standardization. The degree matrix of the instance-level hypergraph; Let be the adjacency matrix of the instance-level hypergraph, representing the connection relationships between nodes in the instance-level hypergraph; For instance-level hypergraphs t The weight matrix of the hyperedge. For bias terms; The propagation coefficients of the instance-level hypergraph; For activation functions; After each round of hypergraph information propagation, the weights of hyperedges in the instance-level hypergraph are adjusted based on node feature similarity, and the formula is updated. , The cosine similarity function; The hypergraph information propagation of modal-level hypergraphs uses the following formula: ； For modal-level hypergraphs t Propagation characteristics of information after propagation in a hypergraph; The propagation coefficients of the modal-level hypergraph; Let be the degree matrix of the modal-level hypergraph; Let be the adjacency matrix of the modal hypergraph, representing the connection relationships between nodes in the modal hypergraph; For modal-level hypergraphs t Wheel hyperedge weight matrix; For bias terms; After each round of hypergraph information propagation, the hyperedge weights in the modal hypergraph are dynamically adjusted based on environmental parameters, using the following formula: ; The environmental impact score is calculated based on ocean current velocity and pressure, with a value range of [0, 0.3]. This is the adjustment coefficient.

4. The pipeline full-state safety assessment method based on multi-dimensional information interconnection and autonomous evolution coordination according to claim 1, characterized in that, The dual-graph information aggregation uses an attention mechanism and feature recombination to fuse the propagation features of instance-level hypergraphs and modality-level hypergraphs, extracting consistency features and specific features. The attention mechanism is a dual-graph attention layer. It calculates the association weights between instance-level hypergraph nodes and modal-level hypergraph nodes, and calculates the cross-graph fusion features of the instance-level and modal-level hypergraphs based on the association weights. The propagation features of the instance-level hypergraph and the cross-graph fusion features are then recombined, and consistent features are output through two branches. and exclusive features These together describe the specific relationship between sensor information and defect information, and are collectively referred to as associated features; The consistency feature The acquisition process is as follows: ； in,[ m ] represents the m-th instance-level hypergraph node. Indicates feature splicing; This represents a one-dimensional convolutional layer; Indicates cross-graph fusion features; The specific features The acquisition process is as follows: in, It is a mean function; The loss function for hypergraph association optimization consists of three parts: propagation loss, etc. fusion loss Mission losses , For classifiers, For real labels, Calculate the cross-entropy; During joint training, the propagation loss coefficient, fusion loss coefficient, and task loss coefficient were set to 0.4, 0.3, and 0.3, respectively.

5. The pipeline full-state safety assessment method based on multi-dimensional information interconnection and autonomous evolution collaboration according to claim 1, characterized in that, The multi-scene feature mining module includes intra-class progressive calibration and inter-class knowledge transfer; The intra-class progressive calibration specifically involves obtaining the current environment state representation vector, extracting the relationship between the current environment state representation vector and associated features using a dual-branch feature extraction structure, and constructing feature consistency constraints based on the extracted relationship. Multi-source environmental parameters, including temperature, are acquired through sensors around the pipeline. T ,pressure P and flow rate V After standardizing the multi-source environmental parameters, a feature fusion algorithm is used to generate a current environmental state representation vector that highly matches the associated features. E The formula is as follows: in, , , The weights of environmental parameters determined based on correlation analysis of pipeline safety assessment Norm ( ) is a standardized function; The dual-branch feature extraction structure includes a global view branch and a local view branch; the global view branch provides consistent features. and exclusive features After positional encoding, the data is processed by a transformer encoder followed by global average pooling to obtain long-range dynamic features, which are then used as the global feature vector F. g Local view branches have consistent features. and exclusive features After performing convolution operations using a 3×1 small convolution kernel, local convolution features are obtained. The local convolution features are then compared with the environmental state representation vector. E The cosine similarity correlation was used, and a gating mechanism was employed to filter local features strongly correlated with environmental perturbations, which were then used as the local feature vector F. l ; the global feature vector F g With local eigenvector F l The features are concatenated to generate a fused feature vector F=[F g F l ]; To align the fused feature vector with the semantic reference features of the samples obtained through pre-training, a semantically aware constrained loss function is designed. To apply feature consistency constraints, the formula is as follows: in, N For the sample size, cos( ) is the cosine similarity function. Let be the fused feature vector of the i-th pipeline defect sample. The sample semantic reference features, which are consistent with the semantics of the i-th pipeline defect sample, are obtained by pre-training on standard environmental data under a stable and invariant environment.

6. The pipeline full-state safety assessment method based on multi-dimensional information interconnection and autonomous evolution coordination according to claim 5, characterized in that, The inter-class knowledge transfer includes spatiotemporal feature clustering and hyperdimensional space constraints; The spatiotemporal feature clustering is as follows; First, spatiotemporal alignment preprocessing is performed on consistency features and specific features to extract temporal and spatial dimension features, and a spatiotemporal feature vector X=[X t X s ], X t X is a subset of time features. s A subset of spatial features; The K-means clustering algorithm is adopted, which measures feature similarity based on cosine similarity and determines the number of clusters K according to the elbow rule to generate multi-class feature representation primitives. The feature representation primitives and the corresponding sample semantic reference features are mapped to a hyperdimensional space, and constrained optimization is performed in the hyperdimensional space. Intragroup compactness constraints Reduce the distance between similar feature representation primitives, C k Let k be the set of feature representation primitives. n k The number of samples in the set. As the class center; constrained by between-group separability Expand the distributional differences of heterogeneous feature representation primitives. The global feature representation primitive center; Introducing spatiotemporal correlation constraints Corr is the environmental evolution rate factor. () represents the correlation coefficient function; Based on the triple constraints of intra-group compactness, inter-group separability, and spatiotemporal correlation, the knowledge of the distribution law of the feature representation primitives of the source class is transferred to the hyperdimensional space of the target domain.

7. The pipeline full-state safety assessment method based on multi-dimensional information interconnection and autonomous evolution collaboration according to claim 1, characterized in that, The pipeline full-state safety assessment model is obtained by training through a total loss function: in 、 、 To constrain the weights, the initial values ​​are all set to 1; First, pre-training is performed using labeled standard environmental data. During pre-training, the multi-scenario feature mining module is turned off to obtain a pre-trained pipeline full-state safety assessment model. Freeze all parameters in the pre-trained evaluation model, unlock the multi-scene feature mining module, and train using target environment data; Using target environment data as training samples, based on the total loss function L total The parameters of the multi-scene feature mining module are updated iteratively using the mini-batch gradient descent algorithm. Each iteration only uses incremental data fragments of the target environment data and does not retrain the entire pre-training pipeline full-state safety assessment model. Set the parameter update threshold to a decrease in the loss function of less than 10. -4 The current update is stopped when the pipeline's full-state safety assessment model parameters are dynamically optimized.

Images