Spatial proximity-awareness method and system for identification of jacket damage and correlation to source

Abstract

This invention belongs to the field of structural health monitoring technology and discloses a spatial proximity sensing method and system for identifying and associating duct stent damage. The invention constructs a physical proximity graph, transforming topological proximity relationships into numerical constraints; extracts bi-branch features from vibration signals and introduces relative position sensing self-attention; constructs category prototypes in a low-dimensional space, calculates logical values ​​based on cosine distance and radius, enhances open set identification using feature mixing and energy scoring, and introduces proximity constraints to maintain feature space continuity; employs threshold discrimination for known / unknown cases; uses linear Warmup and cosine annealing for joint optimization; and for unknown samples, re-ranks them based on similarity and proximity weights, associating them with adjacent reference regions. This invention uses a threshold discrimination strategy, setting differentiated rejection boundaries for different damage categories to avoid over-rejection or mis-acceptance caused by a uniform threshold, thus improving the ability to reject unknown damage and the overall diagnostic reliability.

Description

Technical Field

[0001] This invention belongs to the field of structural health monitoring technology, and in particular relates to a spatial proximity sensing method and system for identifying and associating the source of duct stent damage. Background Technology

[0002] Offshore jacket platforms, as crucial infrastructure for deep-sea oil and gas resource development, operate in complex marine environments for extended periods. With prolonged service, structural components may exhibit damage such as fatigue cracks, loose connections, and localized corrosion thinning. Due to the high safety sensitivity and maintenance costs of jacket structures, failure to promptly identify and address damage can easily lead to the propagation of localized failures or even structural accidents. Therefore, timely and effective damage identification for jacket structures is of significant engineering importance.

[0003] Existing damage identification methods are mainly based on structural dynamics theory. They identify and locate structural damage states by extracting modal parameters such as modal frequencies, mode shapes, and damping ratios, or by analyzing time-domain, frequency-domain, and time-frequency-domain response characteristics. These methods have strong physical interpretability, but they are typically highly dependent on the accuracy of modal parameter extraction and the quality of manual feature design. Furthermore, the modal identification process is easily affected by environmental noise, operating condition fluctuations, and sensor placement conditions. When the structure is in a strongly nonlinear, multi-damage coupled, or time-varying service environment, the sensitivity of the structural modes to single features decreases, thus limiting the damage discrimination capability.

[0004] With the development of deep learning technology, convolutional neural networks have been used to automatically extract damage features from raw responses and perform state classification. However, most existing methods assume that the training and test sets follow the same distribution. Real-world duct structures may encounter novel damage patterns during service that were not seen during the training phase. This leads traditional classifiers to tend to force out-of-distribution samples into the known category with the highest confidence, thus increasing the risk of misclassification.

[0005] Open set identification methods enhance the model's ability to reject unknown samples by constructing energy boundaries, extreme value theories, reconstruction errors, or adversarial discrimination mechanisms. However, most existing methods remain at the level of binary classification between known and unknown, and have not yet effectively answered the engineering question of where unknown damage is more likely to originate. For engineering objects such as jacket structures with clear spatial topological relationships and limited inspection and maintenance resources, simply identifying unknown states is still insufficient to support subsequent decision-making.

[0006] From a structural mechanism perspective, local damage on different sides, at different heights, and on adjacent components of the jacket structure naturally exhibits local correlation and spatial continuity in its dynamic response. If such spatial priors are ignored in the model design, even if high classification accuracy can be achieved under closed-set conditions, it is difficult to guarantee that the discrimination boundary of unknown samples in open-set scenarios has reasonable spatial interpretability.

[0007] Therefore, it is necessary to propose a technical solution that can simultaneously achieve known damage identification, unknown damage rejection, and unknown damage source correlation reasoning.

[0008] Based on the above analysis, the problems and shortcomings of the existing technology are as follows: (1) Limitations of traditional dynamic methods: relying on modal parameter extraction and artificial feature design, the discrimination ability is significantly reduced in strong nonlinear, multi-damage coupling and time-varying marine environments.

[0009] (2) Limitations of deep learning methods: By default, the training set and the test set are distributed in the same way, which makes it impossible to identify novel damage patterns that occur during service. It is easy to misclassify unknown samples as known categories, which may lead to safety risks.

[0010] (3) Shortcomings of existing open set identification methods: They can only distinguish between known and unknown, but cannot answer the engineering decision-making question of where unknown damage is more likely to come from, and cannot provide spatial guidance for inspection and maintenance.

[0011] (4) Structural space priors are not utilized: The local correlation and spatial continuity between different sides, different height layers, and adjacent components of the jacket are ignored by the existing model, resulting in a lack of reasonable spatial interpretation of the unknown sample discrimination boundary in the open set scenario. Summary of the Invention

[0012] To overcome the problems in existing technologies where damage patterns not observed during the training phase are difficult to effectively reject and where unknown damage lacks spatial source correlation capabilities, this invention discloses a spatial proximity sensing method and system for duct stent damage identification and source correlation. The technical solution is as follows: This invention is implemented as follows: a spatial proximity sensing method for identifying and associating the source of duct stent damage includes the following steps: S1. Physical Proximity Graph Construction: The duct structure is abstracted into an undirected graph based on physical spatial adjacency relationships, and the proximity weights between damaged nodes are calculated, transforming the physical topological proximity relationships into numerical constraints. S2. Spatial perception feature extraction: The collected one-dimensional vibration response signal is input into the dual-branch spatial perception feature extraction module. The shallow detail branch extracts local high-frequency disturbance features, and the deep semantic branch introduces a relative position perception self-attention mechanism to extract global high-level features. The local high-frequency disturbance features and global high-level features are then concatenated and mapped to a low-dimensional feature space to obtain a low-dimensional feature representation. S3. Proximity-aware prototype learning: Construct learnable category prototypes and discrimination boundaries in a low-dimensional feature space; apply local continuity constraints in the feature space to physically adjacent damage categories using proximity weights; and use an open set recognition enhancement strategy to distinguish between known and unknown damages. S4. Threshold discrimination: Based on the global optimal threshold and the specific threshold for each predicted category, the sample is classified into a known category or rejected for an unknown category. S5. Model optimization training mechanism: During the training phase, linear Warmup and cosine annealing learning rate scheduling are used to jointly optimize feature extraction, prototype learning, topological constraints and open set boundaries. S6. Unknown damage source association reasoning: For samples determined to be unknown, the candidate known categories are reordered based on their similarity relationship with each known category prototype and the proximity weight. The unknown samples are associated with physically adjacent or locally adjacent reference regions, and the source association results of the unknown damage are output.

[0013] In step S1, the undirected graph is ,in, For the set of nodes of all damage categories, It is a set of adjacent edges in physical space; the jacket is mapped to multiple sides and multiple height layers, and the edge set includes intra-plane adjacency relationships and cross-plane adjacency relationships.

[0014] In step S1, calculating the proximity weights between damaged nodes includes: The breadth-first search algorithm is used to calculate any two damage category nodes. and Shortest path distance between Then the formula for calculating the neighbor weight is: ; In the formula, For neighbor weights, All are damage category nodes. This is the attenuation coefficient.

[0015] In step S2, the shallow detail branch extracts local high-frequency perturbation features by performing convolution operations on receptive fields with kernels smaller than 3; the deep semantic branch gradually expands the receptive field through multi-layer convolution and nonlinear mapping to obtain high-level features with global representation capabilities.

[0016] Furthermore, a relative position-aware self-attention mechanism is introduced into the deep semantic branch, which generates a query matrix through linear mapping for input features. Key matrix Sum matrix ; The elements of the relative position offset matrix are defined as follows: ; In the formula, This is the relative position offset matrix. This refers to the time position index in the input feature sequence. A scale parameter used to control the decay rate; The formula for calculating the attention weight matrix is: ; In the formula, This is the attention weight matrix. It is the transpose symbol. This is the relative position offset matrix. For normalized exponential functions, For feature dimension, For learnable bias scaling parameters; After concatenating the deep features with the high-frequency detail features extracted by the shallow convolution in the channel dimension, the deep features are then mapped to the low-dimensional feature space through the projection head.

[0017] In step S3, constructing a learnable category prototype and discrimination boundary in the low-dimensional feature space includes: Let the prototypes of all known categories be The corresponding prototype radius is Then the sample features With category prototype The cosine distance between them is defined as: ; In the formula, Cosine distance It is an L2 norm; Based on the cosine distance and class-specific radius, the discrimination score of a sample relative to each class is: ; In the formula, The score is the discrimination score of the sample relative to each category. This is the magnification factor. It is a small constant used to prevent the denominator from being zero.

[0018] In step S3, the open set identification enhancement strategy includes: Pseudo-unknown samples are generated through feature mixing. A pseudo-unknown sample is defined as follows: ; In the formula, This is a pseudo-unknown sample. For randomly sampled features of different classes, It is random noise; Energy score is used to measure the confidence that a sample belongs to a known class. The energy score is defined as follows: ; In the formula, To score energy, The total number of known categories. The temperature parameter is used to constrain the energy score of pseudo-unknown samples to be lower than that of true known samples during training, thereby widening the distance between pseudo-unknown samples and known class prototypes.

[0019] In step S4, the threshold determination includes: The global optimal threshold is searched based on the comprehensive score of all test samples, and then the category-specific threshold is searched for each predicted category on the corresponding sample subset. Let the sample The prediction category is The corresponding comprehensive known score is , No. The threshold for the class is The decision rule for unknown samples is: ; In the formula, It is determined to be unknown.

[0020] In step S5, the model optimization training mechanism includes: Let the total number of training rounds be... The Warmup stage length is The initial baseline learning rate is Then the first The learning rate for each training epoch is defined as: ; In the formula, For the first The learning rate for each training round.

[0021] Another object of the present invention is to provide a spatial proximity sensing system for identifying and associating the source of duct stent damage. This system is used to implement the aforementioned spatial proximity sensing method for identifying and associating the source of duct stent damage. The system includes: The physical proximity graph construction module is used to abstract the duct structure into an undirected graph based on physical spatial adjacency relationships, calculate the proximity weights between damaged nodes, and transform the physical topological proximity relationships into numerical constraints. The spatial perception feature extraction module is used to input the collected one-dimensional vibration response signal into the dual-branch spatial perception feature extraction module. The shallow detail branch extracts local high-frequency disturbance features, while the deep semantic branch introduces a relative position perception self-attention mechanism to extract global high-level features. The local high-frequency disturbance features and global high-level features are then concatenated and mapped to a low-dimensional feature space to obtain a low-dimensional feature representation. The proximity-aware prototype learning module is used to construct learnable category prototypes and discrimination boundaries in a low-dimensional feature space; it applies local continuity constraints in the feature space to physically adjacent damage categories using proximity weights, and uses an open set recognition enhancement strategy to distinguish between known and unknown damages. The threshold discrimination module classifies samples into known categories or rejects them into unknown categories based on the globally optimal threshold and the specific threshold for each predicted category. The model optimization training module employs linear Warmup and cosine annealing learning rate scheduling during the training phase to jointly optimize feature extraction, prototype learning, topological constraints, and open set boundaries. The unknown damage source association reasoning module is used to reorder candidate known categories based on the similarity relationship between the unknown sample and each known category prototype, and combined with the proximity weight, to associate the unknown sample with a physically adjacent or locally adjacent reference region, and output the source association result of the unknown damage.

[0022] Combining all the above technical solutions, the beneficial effects of this invention are as follows: First, this invention addresses the problems of identifying unknown damage to duct stent structures and the unclear sources of unknown damage by constructing a spatial proximity sensing framework. This framework first constructs a physical proximity map based on the spatial distribution relationships of different sides and height layers of the duct stent, used to characterize the topological adjacency relationships between different damage categories. Then, vibration response signals are input into a spatial sensing feature extraction module to extract a low-dimensional representation that combines global semantic information with local detailed features. Based on this, a prototype distribution of each known category in the feature space is constructed through proximity sensing prototype learning, and physical proximity constraints are used to enhance the structural correlation between adjacent damage categories. Finally, a threshold strategy is combined to achieve open set discrimination between known and unknown classes, and the source association reasoning from unknown damage to adjacent known regions is completed based on the similarity relationship between unknown samples and known prototypes.

[0023] Secondly, this invention constructs a physical proximity map between damage categories of the ductwork, introducing the spatial topological relationships between different sides, different height layers, and adjacent components into the identification process. This overcomes the problem of neglecting structural spatial priors in existing technologies, improving the consistency between damage identification results and the actual structural distribution. This invention uses a spatial perception feature extraction module to jointly extract deep semantic information and shallow detail information from vibration response signals. This preserves local changes caused by minor damage while enhancing the representation of global response patterns, thereby improving the accuracy and stability of known damage identification. This invention constructs open set discrimination boundaries through proximity-aware prototype learning, pseudo-unknown sample generation, and energy constraints. Compared to existing methods that only apply to closed set classification, this invention can more effectively distinguish novel damage patterns not appearing during the training phase, reducing the risk of unknown damage being misclassified as known damage. This invention employs a threshold discrimination strategy, setting differentiated rejection boundaries for different damage categories. This avoids over-rejection or misacceptance caused by a uniform threshold, thereby improving the ability to reject unknown damage and the overall reliability of diagnosis. After an unknown sample is rejected, this invention further combines prototype similarity and physical proximity to perform source association reasoning. This not only determines whether the sample belongs to unknown damage, but also provides its more likely neighboring reference area, which is beneficial for subsequent manual inspection, fault diagnosis and maintenance decisions. It has better engineering applicability and spatial interpretability.

[0024] Third, this invention enables the effective identification of unknown damage and spatial correlation reasoning of its potential source areas during the service life of jacket structures, thereby significantly reducing the scope of manual inspections and lowering the inspection time and maintenance costs of offshore platforms. Simultaneously, the method of this invention can be embedded into existing structural health monitoring systems to achieve online or near real-time assessment. Most existing structural damage identification methods focus on the classification and identification of known damage categories, or only achieve the rejection of unknown samples within an open set identification framework, but lack the ability to further reason about the source areas of unknown damage. This invention, based on the identification of unknown damage, introduces the physical adjacency relationship of the jacket structure to model the spatial correlation between unknown samples and known damage areas, extending from identifying whether something is unknown to inferring possible sources, demonstrating innovation in combining open set damage identification with spatial correlation analysis. In the field of structural health monitoring, how to simultaneously achieve reliable identification and spatial localization of unknown states under the condition of lacking prior samples of unknown damage has always been a challenging problem. Traditional methods typically can only achieve classification of known categories, or perform unknown rejection under open set conditions, but it is difficult to further determine the potential location of unknown damage. This invention achieves correlation reasoning between unknown damage and known structural areas, alleviating the problem of the inability to locate unknown damage to a certain extent. Existing technologies typically treat unknown samples as a uniform anomaly category, only performing rejection processing without modeling their internal structure or relationships with known categories. This results in a technical approach where the unknown is deemed unanalyzable. This invention, however, goes beyond simply identifying unknown samples by conducting source association reasoning. This approach breaks through the limitations of binary discrimination, extending unknown damage analysis from merely determining whether something is unknown to exploring its possible origins, thus improving the model's spatial interpretability and engineering applicability. Attached Figure Description

[0025] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with the disclosure of this invention and, together with the description, serve to explain the principles of the disclosure of this invention. Figure 1 This is a flowchart of the spatial proximity sensing method for catheter stent damage identification and source association provided in an embodiment of the present invention; Figure 2 This is a roadmap of the spatial proximity sensing method for catheter stent damage identification and source association provided in the embodiments of the present invention; Figure 3 This is a diagram illustrating the overall architecture of the spatial proximity sensing method for catheter stent damage identification and source association provided in this embodiment of the invention. Figure 4 This is a diagram of the spatial perception feature extraction and proximity perception prototype learning structure provided in the embodiments of the present invention; Figure 5 This is a confusion matrix diagram provided in an embodiment of the present invention; Figure 6This is a location-based identification result diagram provided by an embodiment of the present invention. Detailed Implementation

[0026] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Many specific details are set forth in the following description to provide a thorough understanding of the present invention. However, the present invention can be practiced in many other ways different from those described herein, and those skilled in the art can make similar modifications without departing from the spirit of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0027] The innovation of this invention lies in: (1) An unknown damage source association reasoning mechanism is proposed. Based on the unknown damage rejection, the association results between the unknown damage and the known reference area are further output. This is an important innovation that is different from the existing binary classification method that only stays in the known / unknown.

[0028] (2) A spatial proximity perception method for open set damage identification of duct stent structures is proposed, which constructs a physical proximity map of the side distribution and height layer relationship of the duct stent and uses it for damage category modeling.

[0029] (3) A spatial perception feature extraction method combining shallow detail branching, deep semantic branching and relative position perception self-attention mechanism is proposed to improve the feature expression capability of vibration signal.

[0030] (4) A proximity-aware prototype learning mechanism combining category prototype, prototype radius, pseudo-unknown generation and energy constraint is proposed to simultaneously achieve known damage classification and unknown damage rejection.

[0031] (5) A threshold discrimination mechanism is proposed, which sets discrimination thresholds for different prediction categories to improve the discrimination accuracy in open set scenarios.

[0032] Example 1, such as Figure 1 , Figure 2 and Figure 3 As shown, the spatial proximity sensing method for catheter stent damage identification and source association provided in this embodiment of the invention includes the following steps: S1. Physical Proximity Graph Construction: The duct structure is abstracted into an undirected graph based on physical spatial adjacency relationships, and the proximity weights between damaged nodes are calculated, transforming the physical topological proximity relationships into numerical constraints. Abstracting the guide frame structure into an undirected graph is... ,in, For the set of nodes of all damage categories, It is a set of adjacent edges in physical space; the jacket is mapped to multiple sides and multiple height layers, and the edge set includes intra-plane adjacency relationships and cross-plane adjacency relationships.

[0033] The calculation of the proximity weights between damaged nodes includes: The breadth-first search algorithm is used to calculate any two damage category nodes. and Shortest path distance between Then the formula for calculating the neighbor weight is: ; In the formula, For neighbor weights, All are damage category nodes. This represents the attenuation coefficient. This physical proximity metric provides the foundational weighting for prototype constraints and inference reordering in the subsequent feature space. Through this definition, physical topological proximity relationships are transformed into numerical constraints that can participate in feature learning and inference. The closer two damage categories are in the structure, the greater their corresponding weights, thus enabling adjacent damage categories to have a stronger coupling relationship in subsequent prototype learning and source association inference.

[0034] S2. Spatial perception feature extraction: The collected one-dimensional vibration response signal is input into the dual-branch spatial perception feature extraction module. The shallow detail branch extracts local high-frequency disturbance features, and the deep semantic branch introduces a relative position perception self-attention mechanism to extract global high-level features. The local high-frequency disturbance features and global high-level features are then concatenated and mapped to a low-dimensional feature space to obtain a low-dimensional feature representation. The acquired one-dimensional vibration response signal is input into the dual-branch spatial sensing feature extraction module, see Figure 4 The dual-branch module includes shallow detail branches and deep semantic branches.

[0035] The shallow detail branch consists of two one-dimensional convolutional layers with a kernel size of 3 and a stride of 1. The deep semantic branch consists of four one-dimensional convolutional layers with a kernel size of 5 and a stride of 1; the attention mechanism has an L of 32.

[0036] The shallow detail branch extracts local high-frequency perturbation features by performing convolution operations on receptive fields with kernels smaller than 3, in order to preserve fine-grained changes caused by weak damage; the deep semantic branch gradually expands the receptive field through multi-layer convolution and nonlinear mapping to obtain high-level features with stronger global representation capabilities.

[0037] A relative position-aware self-attention mechanism is introduced into the deep semantic branch, which generates a query matrix through linear mapping for input features. Key matrix Sum matrix ; The elements of the relative position offset matrix are defined as follows: ; In the formula, This is the relative position offset matrix. This refers to the time position index in the input feature sequence. The scale parameter controls the decay rate; the closer two positions are on the sequence, the larger their bias value, making it easier to obtain higher weights during attention allocation.

[0038] The formula for calculating the attention weight matrix is: ; In the formula, This is the attention weight matrix. It is the transpose symbol. This is the relative position offset matrix. For normalized exponential functions, For feature dimension, The bias scaling parameter is learnable; through the above calculation, the model can more fully highlight the local structural information related to the damage-sensitive area while modeling global dependencies.

[0039] After concatenating the deep features with the high-frequency detail features extracted by the shallow convolution in the channel dimension, the deep features are then mapped to the low-dimensional feature space through the projection head.

[0040] S3. Proximity-aware prototype learning: Construct learnable category prototypes and discrimination boundaries in a low-dimensional feature space; apply local continuity constraints in the feature space to physically adjacent damage categories using proximity weights; and use an open set recognition enhancement strategy to distinguish between known and unknown damages. To achieve accurate classification of known damage and effective rejection of unknown damage, learnable category prototypes are constructed in a low-dimensional feature space, such as... Figure 4 As shown. Let the prototypes of all known categories be... The corresponding prototype radius is Then the sample features With category prototype The cosine distance between them is defined as: ; In the formula, Cosine distance It is an L2 norm; Based on the cosine distance and class-specific radius, the discrimination score of a sample relative to each class is: ; In the formula, The score is the discrimination score of the sample relative to each category. This is the magnification factor. This is a small constant used to prevent the denominator from being zero. Through the above dynamic boundary construction method, different categories can form closed discrimination regions of different scales, thereby adapting to the discreteness of the feature distribution of different damage categories.

[0041] To enhance the model's ability to reject unknown damage, pseudo-unknown samples are generated through feature mixing; preferably, pseudo-unknown samples are defined as: ; In the formula, This is a pseudo-unknown sample. For randomly sampled features of different classes, It is random noise; Energy score is used to measure the confidence that a sample belongs to a known class. The energy score is defined as follows: ; In the formula, To score energy, The total number of known categories. The temperature parameter is used to constrain the energy score of pseudo-unknown samples to be lower than that of real known samples during training, and to increase the distance between pseudo-unknown samples and known class prototypes, thereby improving the boundary perception capability of unknown working conditions.

[0042] Furthermore, the neighbor weights obtained in step S1 are introduced into the prototype learning process to ensure that physically adjacent damage categories maintain reasonable local continuity in the feature space, thereby enhancing the spatial consistency and interpretability of the feature space.

[0043] S4. Threshold discrimination: Based on the global optimal threshold and the specific threshold for each predicted category, the sample is classified into a known category or rejected for an unknown category. Unknown sample identification is achieved through a threshold strategy. Preferably, a globally optimal threshold is first searched based on the comprehensive score of all test samples, and then a category-specific threshold is searched for each predicted category on its corresponding sample subset. The comprehensive score is an energy score calculated based on an energy function.

[0044] Let the sample The prediction category is The corresponding comprehensive known score is , No. The threshold for the class is The decision rule for unknown samples is: ; In the formula, To determine whether something is unknown, different rejection criteria can be used for different known categories through the above-mentioned decision-making method, thereby avoiding the problem of over-rejection or mis-acceptance caused by a uniform threshold.

[0045] S5. Model optimization training mechanism: During the training phase, linear Warmup and cosine annealing learning rate scheduling are used to jointly optimize feature extraction, prototype learning, topological constraints and open set boundaries. Since this invention simultaneously optimizes feature extraction, prototype learning, topological constraints, and open set boundaries, the coupling of multiple losses can increase instability in the early stages of training. Therefore, a linear Warmup and cosine annealing learning rate scheduling method is used during the training phase.

[0046] The model optimization training mechanism includes: assuming the total number of training rounds is... The Warmup stage length is The initial baseline learning rate is Then the first The learning rate for each training epoch is defined as: ; In the formula, For the first The learning rate for each training epoch. Through the above scheduling method, the learning rate is gradually increased in the early stage of training to enable the network to establish stable basic feature representations first; after the warmup, the learning rate is gradually decreased so that the model can perform smoother and more refined joint optimization of class boundaries, prototype distribution and proximity constraints in the later stage.

[0047] S6. Unknown damage source association reasoning: For samples determined to be unknown, the candidate known categories are reordered based on their similarity relationship with each known category prototype and the proximity weight. The unknown samples are associated with physically adjacent or locally adjacent reference regions, and the source association results of the unknown damage are output.

[0048] For a sample x determined to be unknown, its low-dimensional feature representation is extracted, and its similarity to each known category prototype is calculated. This similarity is multiplied by the physical proximity weight of the corresponding category to obtain the association score of the unknown sample relative to the i-th known category: where represents the association score between the unknown sample and the i-th known category, represents the similarity between the features of the unknown sample and the category prototype, and represents the proximity weight determined by physical topology. The candidate known categories are reordered from largest to smallest, and the regions corresponding to the top-ranked known categories are identified as the source association regions of the unknown damage.

[0049] The correlation results are used to provide a reference for manual inspection and maintenance priority planning, enabling the model to not only answer whether a sample belongs to an unknown state, but also to answer which known reference area the unknown state is more likely to correspond to.

[0050] Example 2: The spatial proximity sensing system for catheter stent damage identification and source association provided in this embodiment of the invention includes: The physical proximity graph construction module is used to abstract the duct structure into an undirected graph based on physical spatial adjacency relationships, calculate the proximity weights between damaged nodes, and transform the physical topological proximity relationships into numerical constraints. The spatial perception feature extraction module is used to input the collected one-dimensional vibration response signal into the dual-branch spatial perception feature extraction module. The shallow detail branch extracts local high-frequency disturbance features, while the deep semantic branch introduces a relative position perception self-attention mechanism to extract global high-level features. The local high-frequency disturbance features and global high-level features are then concatenated and mapped to a low-dimensional feature space to obtain a low-dimensional feature representation. The proximity-aware prototype learning module is used to construct learnable category prototypes and discrimination boundaries in a low-dimensional feature space; it applies local continuity constraints in the feature space to physically adjacent damage categories using proximity weights, and uses an open set recognition enhancement strategy to distinguish between known and unknown damages. The threshold discrimination module classifies samples into known categories or rejects them into unknown categories based on the globally optimal threshold and the specific threshold for each predicted category. The model optimization training module employs linear Warmup and cosine annealing learning rate scheduling during the training phase to jointly optimize feature extraction, prototype learning, topological constraints, and open set boundaries. The unknown damage source association reasoning module is used to reorder candidate known categories based on the similarity relationship between the unknown sample and each known category prototype, and combined with the proximity weight, to associate the unknown sample with a physically adjacent or locally adjacent reference region, and output the source association result of the unknown damage.

[0051] To further demonstrate the positive effects of the above embodiments, the present invention conducts the following experiments based on the above technical solution: Experimental research is carried out based on a scaled-down model of a marine jacket. Flange connection structures are introduced at the diagonal and vertical bracing components of the jacket, and adjacent flanges are connected by bolts. By adjusting the bolt tightness and removing bolts, local connection stiffness degradation or failure is simulated, resulting in 16 different damage scenarios. Structural excitation is generated by hammer impact, and structural vibration response signals are synchronously collected using an accelerometer. For each damage location, 20 hammer impact tests are repeated. The resulting jacket structure damage identification dataset contains 16 different structural damage conditions, with 20 samples in each category, and each sample consisting of 2048 time-series sampling points. Twelve categories are randomly selected as known categories for training and testing, while the remaining four categories are unknown and completely invisible during the training phase, only introduced during the testing phase. The known category samples are divided into training and testing sets at a ratio of 80% and 20%, respectively.

[0052] Figure 5 The confusion matrix consisting of all known classes and unknown samples is presented. The results show that most known classes are stably and accurately identified. Figure 6 As shown, unknown damage classes 1, 11, and 16 are highly concentrated among known damage classes 13, 12, and 15 (accounting for 90%, 95%, and 80%, respectively), while known damage class 6 is mainly distributed between known damage classes 7 and 5. This non-random concentration mapping indicates that the model's determination of unknown sources is based on information from learned prototype structures, rather than random guessing. The high concentration of unknown samples towards a single known location not only reflects the consistency of their structural response patterns but also confirms the proximity of their physical locations. The dispersed distribution towards adjacent known classes suggests that the unknown damage is located within the transition boundary or local neighborhood spanned by the known prototypes, which is consistent with the adjacent physical structures of known damage classes 5, 6, and 7. This demonstrates that the model can capture the spatial topological relationships of damage locations, thus exhibiting significant robustness and physical plausibility in feature space modeling.

[0053] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any modifications, equivalent substitutions and improvements made by those skilled in the art within the scope of the technology disclosed in the present invention and within the spirit and principles of the present invention should be covered within the scope of protection of the present invention.

Claims

1. A spatial proximity sensing method for identifying and associating the source of duct stent damage, characterized in that, The method includes the following steps: S1. Physical Proximity Graph Construction: The duct structure is abstracted into an undirected graph based on physical spatial adjacency relationships, and the proximity weights between damaged nodes are calculated, transforming the physical topological proximity relationships into numerical constraints. S2. Spatial perception feature extraction: The collected one-dimensional vibration response signal is input into the dual-branch spatial perception feature extraction module. The shallow detail branch extracts local high-frequency disturbance features, and the deep semantic branch introduces a relative position perception self-attention mechanism to extract global high-level features. The local high-frequency disturbance features and global high-level features are then concatenated and mapped to a low-dimensional feature space to obtain a low-dimensional feature representation. S3. Proximity-aware prototype learning: Construct learnable category prototypes and discrimination boundaries in a low-dimensional feature space; apply local continuity constraints in the feature space to physically adjacent damage categories using proximity weights; and use an open set recognition enhancement strategy to distinguish between known and unknown damages. S4. Threshold discrimination: Based on the global optimal threshold and the specific threshold for each predicted category, the sample is classified into a known category or rejected for an unknown category. S5. Model optimization training mechanism: During the training phase, linear Warmup and cosine annealing learning rate scheduling are used to jointly optimize feature extraction, prototype learning, topological constraints and open set boundaries. S6. Unknown damage source association reasoning: For samples determined to be unknown, the candidate known categories are reordered based on their similarity relationship with each known category prototype and the proximity weight. The unknown samples are associated with physically adjacent or locally adjacent reference regions, and the source association results of the unknown damage are output.

2. The spatial proximity sensing method for duct stent damage identification and source association according to claim 1, characterized in that, In step S1, the undirected graph is ,in, For the set of nodes of all damage categories, It is a set of adjacent edges in physical space; the jacket is mapped to multiple sides and multiple height layers, and the edge set includes intra-plane adjacency relationships and cross-plane adjacency relationships.

3. The spatial proximity sensing method for identifying and associating the source of duct stent damage according to claim 1, characterized in that, In step S1, calculating the proximity weights between damaged nodes includes: The breadth-first search algorithm is used to calculate any two damage category nodes. and Shortest path distance between Then the formula for calculating the neighbor weight is: ； In the formula, For neighbor weights, All are damage category nodes. This is the attenuation coefficient.

4. The spatial proximity sensing method for identifying and associating the source of duct stent damage according to claim 1, characterized in that, In step S2, the shallow detail branch extracts local high-frequency perturbation features by performing convolution operations on receptive fields with kernels smaller than 3; the deep semantic branch gradually expands the receptive field through multi-layer convolution and nonlinear mapping to obtain high-level features with global representation capabilities.

5. The spatial proximity sensing method for duct stent damage identification and source association according to claim 4, characterized in that, A relative position-aware self-attention mechanism is introduced into the deep semantic branch, which generates a query matrix through linear mapping for input features. Key matrix Sum matrix ; The elements of the relative position offset matrix are defined as follows: ； In the formula, This is the relative position offset matrix. This refers to the time position index in the input feature sequence. A scale parameter used to control the decay rate; The formula for calculating the attention weight matrix is: ； In the formula, This is the attention weight matrix. It is the transpose symbol. This is the relative position offset matrix. For normalized exponential functions, For feature dimension, For learnable bias scaling parameters; After concatenating the deep features with the high-frequency detail features extracted by the shallow convolution in the channel dimension, the deep features are then mapped to the low-dimensional feature space through the projection head.

6. The spatial proximity sensing method for identifying and associating the source of duct stent damage according to claim 1, characterized in that, In step S3, constructing a learnable category prototype and discrimination boundary in the low-dimensional feature space includes: Let the prototypes of all known categories be The corresponding prototype radius is Then the sample features With category prototype The cosine distance between them is defined as: ； In the formula, Cosine distance It is an L2 norm; Based on the cosine distance and class-specific radius, the discrimination score of a sample relative to each class is: ； In the formula, The score is the discrimination score of the sample relative to each category. This is the magnification factor. It is a small constant used to prevent the denominator from being zero.

7. The spatial proximity sensing method for duct stent damage identification and source association according to claim 6, characterized in that, In step S3, the open set identification enhancement strategy includes: Pseudo-unknown samples are generated through feature mixing. A pseudo-unknown sample is defined as follows: ； In the formula, This is a pseudo-unknown sample. For randomly sampled features of different classes, It is random noise; Energy score is used to measure the confidence that a sample belongs to a known class. The energy score is defined as follows: ； In the formula, To score energy, The total number of known categories. The temperature parameter is used to constrain the energy score of pseudo-unknown samples to be lower than that of true known samples during training, thereby widening the distance between pseudo-unknown samples and known class prototypes.

8. The spatial proximity sensing method for identifying and associating the source of duct stent damage according to claim 1, characterized in that, In step S4, the threshold determination includes: The global optimal threshold is searched based on the comprehensive score of all test samples, and then the category-specific threshold is searched for each predicted category on the corresponding sample subset. Let the sample The prediction category is The corresponding comprehensive known score is , No. The threshold for the class is The decision rule for unknown samples is: ； In the formula, It is determined to be unknown.

9. The spatial proximity sensing method for identifying and associating the source of duct stent damage according to claim 1, characterized in that, In step S5, the model optimization training mechanism includes: Let the total number of training rounds be... The Warmup stage length is The initial baseline learning rate is Then the first The learning rate for each training epoch is defined as: ； In the formula, For the first The learning rate for each training round.

10. A spatial proximity sensing system for identifying and associating the source of duct stent damage, characterized in that, This system is used to implement the spatial proximity sensing method for duct stent damage identification and source association as described in any one of claims 1 to 9, the system comprising: The physical proximity graph construction module is used to abstract the duct structure into an undirected graph based on physical spatial adjacency relationships, calculate the proximity weights between damaged nodes, and transform the physical topological proximity relationships into numerical constraints. The spatial perception feature extraction module is used to input the collected one-dimensional vibration response signal into the dual-branch spatial perception feature extraction module. The shallow detail branch extracts local high-frequency disturbance features, while the deep semantic branch introduces a relative position perception self-attention mechanism to extract global high-level features. The local high-frequency disturbance features and global high-level features are then concatenated and mapped to a low-dimensional feature space to obtain a low-dimensional feature representation. The proximity-aware prototype learning module is used to construct learnable category prototypes and discrimination boundaries in a low-dimensional feature space; it applies local continuity constraints in the feature space to physically adjacent damage categories using proximity weights, and uses an open set recognition enhancement strategy to distinguish between known and unknown damages. The threshold discrimination module classifies samples into known categories or rejects them into unknown categories based on the globally optimal threshold and the specific threshold for each predicted category. The model optimization training module employs linear Warmup and cosine annealing learning rate scheduling during the training phase to jointly optimize feature extraction, prototype learning, topological constraints, and open set boundaries. The unknown damage source association reasoning module is used to reorder candidate known categories based on the similarity relationship between the unknown sample and each known category prototype, and combined with the proximity weight, to associate the unknown sample with a physically adjacent or locally adjacent reference region, and output the source association result of the unknown damage.

Images