A consensus divergence path alignment welding visual implicit knowledge graph construction method and device

Abstract

The application discloses a welding visual implicit knowledge graph construction method and device based on consensus divergence path alignment, and belongs to the technical field of industrial quality detection and artificial intelligence. The method comprises the following steps: constructing a directed cognitive subgraph, establishing a node initial representation and performing relationship perception information propagation to obtain a relationship perception node representation; calculating a matching cost of a cognitive node, performing one-to-one matching and multi-expert consistency checking, calculating expert coverage to determine a consensus candidate concept, projecting each expert cognitive chain to a candidate cognitive concept space; comparing the sequence relationship of the consensus candidate concept in the common expert cognitive chain to determine the stable sequence relationship across experts, establishing a sequence directed graph, and extracting a consensus cognitive path; determining a consensus anchor point corresponding to the consensus cognitive path in the expert cognitive chain, forming a divergence path and its left and right boundaries, and constructing a welding visual implicit knowledge graph. The application realizes the unification of multi-expert cognitive concepts, the extraction of consensus cognitive paths and divergence paths, and the construction of a graph.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and industrial quality inspection technology, and in particular to a method and apparatus for constructing a visual implicit knowledge graph for welding based on consensus divergence path alignment. Background Technology

[0002] Existing welding vision inspection systems typically acquire weld images and use image processing, feature extraction, or deep learning models to identify typical defects such as cracks, porosity, undercut, and lack of fusion. For standardized defects with stable visual features, clear defect types, and sufficient training samples, existing methods have relatively stable recognition capabilities. However, in actual welding inspections, there are still atypical defects with blurred boundaries, weakened features, similar morphologies, or low frequency of occurrence. Simply relying on preset rules and data-driven models is often insufficient for accurate identification, and in such cases, comprehensive judgment by welding inspection experts with rich practical experience is still required.

[0003] Existing research has established a certain technical foundation in multi-source knowledge alignment, consensus process extraction, and divergent knowledge representation. In multi-source knowledge alignment, methods such as semantic representation, entity alignment, and graph matching primarily rely on node attributes, semantic similarity, and local structure to establish cross-source correspondences. However, the same visual cognitive concept may have different expressions and relational roles in different experts' cognitive chains, and the correspondences between multiple experts need to maintain overall consistency. Simply relying on pairwise local matching is insufficient to fully characterize the global concept correspondence among multiple experts. In consensus cognition extraction, sequence pattern mining, multi-sequence matching, and graph representation learning typically identify repetitive structures based on node co-occurrence, sequence overlap, or topological similarity. However, during expert visual inspection, behaviors such as backtracking, local reordering, and path jumping occur, meaning that paths with the same cognitive meaning may not be completely consistent on the surface sequence. Utilizing only node or local sequence similarity is insufficient to simultaneously maintain multi-expert source coverage, cognitive order, and path continuity.

[0004] Furthermore, while existing multi-source knowledge fusion and heterogeneous graph modeling methods can preserve the differences among experts through source labels, node types, and relationship types, they still fall short in expressing the structural relationship between consensus and divergent cognition. Disagreements in welding visual inspection typically do not arise independently but are triggered by abnormal visual cues at specific points in the shared cognitive process. Their starting point, scope, and semantic meaning are closely related to the preceding and following cognitive paths. If only divergent nodes or relationships are recorded as independent information, it is difficult to accurately determine the consensus cognitive position corresponding to the divergent path, and it is also difficult to form a clearly structured relationship between consensus and divergent knowledge.

[0005] Existing technologies have been able to make explicit, structured, and path-based expressions of tacit knowledge of expert behavior, but there is still a lack of technical solutions for aligning the ordered visual cognitive chains of multiple experts at the path level for welding visual inspection scenarios, and further distinguishing consensus cognitive paths, shared path segments, and divergent paths. There is also a lack of methods to unify consensus knowledge and expert-specific divergent knowledge into a welding visual tacit knowledge graph.

[0006] Existing multi-source knowledge alignment methods struggle to establish stable and consistent cross-expert concept correspondences when processing the visual cognitive knowledge of multiple welding inspection experts. Current semantic representation, entity alignment, and graph matching methods typically perform pairwise matching based on node attributes, semantic similarity, or local structure. However, the same visual cognitive concept may have different representations, connections, and cognitive roles in the cognitive chains of different experts. Furthermore, inconsistencies in correspondences may arise when local matching results are extended to multiple expert sources. As the number of experts and cognitive nodes increases, the number of candidate nodes and relationship combinations to be compared grows rapidly, easily leading to increased computational load for matching, time-consuming manual verification, and decreased concept alignment efficiency, ultimately affecting the unified determination of nodes and relationships in the subsequent graph.

[0007] Existing consensus cognition extraction methods struggle to accurately identify complete cognitive processes with stable consistency among multiple experts. Sequence pattern mining, multiple sequence matching, and graph representation learning typically identify recurring structures through node co-occurrence, sequence overlap, or topological similarity. However, during welding visual inspection, behaviors such as backtracking, changes in local observation order, and cross-regional jumps are common, causing the cognitive chains formed when different experts express the same cognitive process to be inconsistent. Judging solely based on node repetition or local sequence similarity can easily lead to the identification of coincidentally overlapping cognitive fragments as consensus, and may also miss stable paths with the same cognitive meaning but with local order perturbations. This necessitates additional processes of repetition comparison and manual verification, thereby increasing the time cost and processing complexity of consensus cognition path extraction.

[0008] Existing methods for representing divergent knowledge struggle to accurately describe the structural relationship between expert-discretionary cognition and consensus cognition. While multi-source knowledge fusion and heterogeneous graph modeling can preserve the differences among experts through source labels and node types, divergences in welding visual inspection are typically not independent but rather arise at specific points along a shared cognitive path triggered by visual cues. Their starting and ending points, as well as their semantic scope, are all related to the preceding and following cognitive processes. If only divergent nodes or relationships are preserved as independent knowledge, it can easily lead to unclear divergence boundaries, ambiguous consensus cognitive paths, and mixed relationships between different divergent paths. This makes it difficult for the constructed graph to simultaneously and clearly express both consensus cognitive paths and expert-discretionary paths, and increases the complexity of graph structure organization and relationship determination. Summary of the Invention

[0009] The purpose of this invention is to overcome the shortcomings of the prior art and provide a method and apparatus for constructing a welding visual implicit knowledge graph with consensus divergence path alignment, which solves the problems of unstable concept correspondence between visual cognitive knowledge of multiple welding inspection experts, difficulty in stably identifying consensus cognitive paths in multi-expert visual cognitive chains, and unclear structural attribution between divergence cognition and consensus cognition.

[0010] To achieve the above objectives, the present invention is implemented using the following technical solution:

[0011] This invention provides a method for constructing a welding visual implicit knowledge graph based on consensus divergence path alignment, comprising:

[0012] Obtain visual cognitive records from several experts regarding the same welding visual inspection object, and construct a directed cognitive subgraph;

[0013] For nodes in a directed cognitive subgraph, an initial node representation is established. Then, relation-aware information is propagated on the relations in the directed cognitive subgraph to obtain relation-aware node representations.

[0014] The matching cost is calculated for the cognitive nodes represented by the relation-aware nodes. The cross-expert node correspondence is determined by one-to-one matching. The cross-expert node correspondence is checked for multi-expert consistency to obtain candidate cognitive concepts and the mapping relationship from cognitive nodes to candidate cognitive concepts.

[0015] The expert coverage is calculated based on the occurrence of candidate cognitive concepts in each expert cognitive chain. Consensus candidate concepts are determined based on the expert coverage and a preset consensus threshold. Based on the mapping relationship between cognitive nodes and candidate cognitive concepts, each expert cognitive chain is projected into the candidate cognitive concept space.

[0016] In the common expert cognitive chain, the order relationship between consensus candidate concepts is compared. Based on the order consistency rate and the preset order threshold, the cross-expert stable order relationship is determined, an order directed graph is established, and continuous directed paths that satisfy the expert coverage constraint and cognitive order constraint in the order directed graph are extracted to obtain the consensus cognitive path.

[0017] Consensus anchors corresponding to consensus cognitive paths in the expert cognitive chain are determined by sequential consistency matching. Continuous cognitive segments that are not absorbed by the consensus cognitive path between adjacent consensus anchors are extracted to form divergent paths and their left and right boundaries. Based on the consensus cognitive path, divergent paths and their left and right boundaries, and expert visual context relationships, a welding visual implicit knowledge graph is constructed.

[0018] Optionally, the visual cognition record includes at least the visual attention area, the order of attention, the corresponding cognition or judgment content, and the visual cues that trigger the cognition.

[0019] The nodes in the directed cognitive subgraph include cognitive nodes and context nodes, and the relationships include at least cognitive advancement relationships and context-driven relationships.

[0020] Optionally, the initial representation of the node is as follows:

[0021] ;

[0022] in, Represents a node The initial representation of the nodes; Indicates a node The semantic vector is obtained by text encoding the name and description; Represents a node Node type; Represents a node Normalized in-degree; Represents a node Normalized degree; Represents a node Topological relative positions in a directed cognitive subgraph; Represents a node The relative order within the expert cognitive chain; This indicates vector concatenation;

[0023] The relation-aware node representation is as follows:

[0024] ;

[0025] , ;

[0026] in, Representing nodes respectively In the Layer, First Layer-aware node representation; Indicates the first The state transformation matrix of the nodes in the layer; Represents a set of relation types; Representing relations In the Layered nodes Aggregated neighborhood information; Represents a nonlinear function; Represents a node Relation-aware node representation at level 0; Represents a node In relation The set of neighbors below; Indicates the first Hierarchical Relationship The transformation matrix; Represents a node In the Layer-aware node representation; This indicates taking the absolute value.

[0027] Optionally, the matching cost is expressed as:

[0028] ;

[0029] , , ;

[0030] in, Experts Corresponding cognitive nodes With experts Corresponding cognitive nodes The matching cost between them; They represent experts respectively. Corresponding cognitive nodes With experts Corresponding cognitive nodes Semantic differences, sequence differences, and local structural differences between them; All of these represent non-negative weights; They represent experts respectively. Corresponding cognitive nodes semantic representation vectors, experts Corresponding cognitive nodes The semantic representation vector; They represent experts respectively. Corresponding cognitive nodes The order position value, expert Corresponding cognitive nodes The sequential position value; They represent experts respectively. Corresponding cognitive nodes Local structure representation vector, expert Corresponding cognitive nodes The local structure representation vector; Indicates taking the absolute value; Represents the L2 norm;

[0031] The cross-expert node correspondence is represented as follows:

[0032] ;

[0033] in, Experts Corresponding cognitive nodes and experts The corresponding cross-expert node correspondence between cognitive nodes; Experts Corresponding cognitive nodes With experts Corresponding cognitive nodes Matching indicator variables between them; This represents the candidate matching matrix corresponding to the cross-expert node correspondence that satisfies the preset matching constraints. In the process, determine the matching matrix that minimizes the total matching cost;

[0034] A multi-expert consistency check is performed on the cross-expert node correspondence to obtain candidate cognitive concepts and the mapping relationship from cognitive nodes to candidate concepts, including:

[0035] If experts Corresponding cognitive nodes and experts If there is a cross-expert node correspondence between corresponding cognitive nodes, and the two form a consistent cross-source correspondence through the node corresponding to another expert, then the cross-expert node correspondence is retained to obtain the candidate cognitive concept and the mapping relationship from cognitive node to candidate concept.

[0036] If multiple cross-expert node correspondences exist, the cross-expert node correspondences with the lowest matching cost and the highest consistency are retained to obtain candidate cognitive concepts and the mapping relationship from cognitive nodes to candidate concepts.

[0037] The candidate cognitive concepts are represented as follows:

[0038] ;

[0039] in, Represents a set of candidate cognitive concepts; They represent the 1st, 2nd, and 3rd respectively. One candidate cognitive concept;

[0040] The mapping relationship from the cognitive node to the candidate cognitive concept is represented as follows:

[0041] ;

[0042] in, Representing cognitive nodes The mapping relationship to candidate cognitive concepts; Indicates the first One candidate cognitive concept.

[0043] Optionally, the expert coverage is expressed as:

[0044] ;

[0045] in, Indicates the first Expert coverage corresponding to each candidate cognitive concept; Indicates containing the first A set of experts for each candidate cognitive concept; Represents a set of experts; Indicates taking the absolute value;

[0046] The consensus candidate concept is represented as follows:

[0047] ;

[0048] in, Represents the set of consensus candidate concepts; Indicates the first One consensus candidate concept; Represents a set of candidate cognitive concepts; This indicates a preset consensus threshold;

[0049] The expert cognitive chains projected onto the candidate cognitive concept space are represented as follows:

[0050] ;

[0051] in, Experts The cognitive chain; Experts The first in the cognitive chain A cognitive node; Experts The number of cognitive nodes in the cognitive chain; Experts The first in the cognitive chain The mapping relationship between cognitive nodes and candidate cognitive concepts.

[0052] Optionally, the order consistency rate is expressed as:

[0053] ;

[0054] in, Indicates the first The consensus candidate concept and the first The order consistency rate among consensus candidate concepts; Indicates containing the first The consensus candidate concept and the first A common set of experts for a consensus candidate concept; They represent the first Consensus candidate concepts , No. Consensus candidate concepts In experts The position where it first appears in the cognitive chain; Indicates an indicator function; Indicates taking the absolute value;

[0055] The cross-expert stable order relationship is represented as follows:

[0056] ;

[0057] in, Indicates a stable order relationship across experts; Indicates the preset order threshold;

[0058] The ordered directed graph is represented as follows:

[0059] ;

[0060] in, Represents a directed graph in sequence; This represents the set of consensus candidate concepts.

[0061] Optionally, continuous directed paths satisfying expert coverage constraints and cognitive order constraints are extracted from the sequential directed graph to obtain consensus cognitive paths, including:

[0062] Using dynamic programming, the path length of the consensus candidate concept in the sequential directed graph is calculated. The consensus candidate concept corresponding to the longest path length is selected and backtracked to obtain the consensus cognition path.

[0063] The path length is expressed as:

[0064] ;

[0065] in, They respectively represent the first Consensus candidate concepts , No. Consensus candidate concepts The path length to the destination; Indicates the first Consensus candidate concepts The set of predecessors; This indicates taking the maximum value;

[0066] The consensus cognition path is represented as follows:

[0067] ;

[0068] in, This represents a consensus-based cognitive path; These respectively represent the first consensus candidate concept, the second consensus candidate concept, and the third consensus candidate concept in the consensus cognition path. One consensus candidate concept; This indicates the length of the consensus cognition path.

[0069] Optionally, consensus anchor points corresponding to the consensus cognition path in the expert cognition chain are determined through sequential consistency matching, including:

[0070] Calculate the maximum sequential matching length between candidate cognitive concepts in the expert cognitive chain and consensus candidate concepts in the consensus cognitive path, and perform reverse backtracking based on the maximum sequential matching length to determine the consensus anchor point in the expert cognitive chain corresponding to the consensus cognitive path.

[0071] The maximum order matching length between candidate cognitive concepts in the expert cognitive chain and consensus candidate concepts in the consensus cognitive path is expressed as:

[0072] ;

[0073] ;

[0074] in, Experts In the cognitive chain Candidate cognitive concepts mapped from each cognitive node and the consensus cognitive path (previous) The maximum order matching length among consensus candidate concepts; Experts In the cognitive chain Candidate cognitive concepts mapped from each cognitive node and the consensus cognitive path (previous) The maximum order matching length among consensus candidate concepts; Experts In the cognitive chain Candidate cognitive concepts mapped from each cognitive node and the consensus cognitive path (previous) The maximum order matching length among consensus candidate concepts; Experts In the cognitive chain Candidate cognitive concepts mapped from each cognitive node and the consensus cognitive path (previous) The maximum order matching length among consensus candidate concepts; Experts The first in the cognitive chain A cognitive node; The first step in the consensus cognition path One consensus candidate concept; This indicates taking the maximum value;

[0075] The branching path is represented as follows:

[0076] ;

[0077] in, Experts The corresponding number A divergent path; Experts The first in the cognitive chain A cognitive node; They represent experts respectively. The corresponding number The starting and ending positions of the diverging paths in the cognitive chain;

[0078] The left boundary of the divergence path is the consensus anchor point position corresponding to the left side of the divergence path, and the right boundary of the divergence path is the consensus anchor point position corresponding to the right side of the divergence path. The divergence path located between two consensus anchor points has left and right boundaries. The divergence path occurring before the first cognitive node in the consensus cognitive path has a right boundary. The divergence path occurring after the last cognitive node in the consensus cognitive path and not re-entering the subsequent consensus cognitive process has a left boundary.

[0079] Optionally, based on consensus-based cognitive paths, divergent paths and their left and right boundaries, and expert visual contextual relationships, a welding visual implicit knowledge graph is constructed, including:

[0080] When a divergent path is generated after a cognitive node in the consensus cognitive path, a divergent trigger edge is formed by pointing from the cognitive node in the consensus cognitive path to the cognitive node at the beginning of the divergent path. The divergent path edges are formed inside the divergent path according to the order of expert cognition.

[0081] When a divergent path re-enters the subsequent consensus cognition process, a divergent return edge is formed by the cognitive node at the end of the divergent path pointing to the cognitive node in the subsequent consensus cognition path.

[0082] The cognitive nodes in the consensus cognitive path are connected through cognitive advancement relationships according to the cognitive order of the consensus cognitive path. The divergent cognitive path is generated from the cognitive nodes in the consensus cognitive path through divergence triggering edges, and the cognitive order is maintained through divergence path edges. When re-entering the consensus cognitive process, it is connected to the cognitive nodes in the subsequent consensus cognitive path through divergence return edges.

[0083] By preserving the context nodes and context-driven relationships in the directed cognitive subgraph, a well-constructed visual implicit knowledge graph of welding is obtained.

[0084] On the other hand, the present invention provides a welding visual implicit knowledge graph construction device for consensus divergence path alignment, comprising:

[0085] The subgraph construction module is used to: acquire visual cognitive records formed by several experts on the same welding visual inspection object, and construct a directed cognitive subgraph;

[0086] The node representation module is used to: establish initial node representations for nodes in a directed cognitive subgraph, propagate relation-aware information on the initial node representations in the relation of the directed cognitive subgraph, and obtain relation-aware node representations;

[0087] The matching and checking module is used to: calculate the matching cost for cognitive nodes represented by relation-aware nodes, determine the cross-expert node correspondence through one-to-one matching, perform multi-expert consistency checks on the cross-expert node correspondence, and obtain candidate cognitive concepts and the mapping relationship from cognitive nodes to candidate cognitive concepts.

[0088] The concept mapping module is used to: calculate expert coverage based on the occurrence of candidate cognitive concepts in each expert cognitive chain; determine consensus candidate concepts based on expert coverage and a preset consensus threshold; and project each expert cognitive chain to the candidate cognitive concept space based on the mapping relationship from cognitive nodes to candidate cognitive concepts.

[0089] The consensus extraction module is used to: compare the order of consensus candidate concepts in the common expert cognitive chain, determine the cross-expert stable order relationship based on the order consistency rate and the preset order threshold, establish an ordered directed graph, extract the continuous directed paths in the ordered directed graph that satisfy the expert coverage constraint and the cognitive order constraint, and obtain the consensus cognitive path;

[0090] The graph construction module is used to: determine the consensus anchor points in the expert cognitive chain corresponding to the consensus cognitive path through sequential consistency matching; extract continuous cognitive segments between adjacent consensus anchor points that are not absorbed by the consensus cognitive path to form divergent paths and their left and right boundaries; and construct a welding visual implicit knowledge graph based on the consensus cognitive path, divergent paths and their left and right boundaries, and expert visual context relationships.

[0091] Compared with the prior art, the beneficial effects achieved by the present invention are as follows:

[0092] This invention unifies multi-expert candidate cognitive concepts by combining node semantics, cognitive order, and local relational structure. This reduces the erroneous merging and separation of concepts caused by matching based solely on text semantics or local neighborhoods, and improves the consistency of cross-expert cognitive node correspondences. It organizes experts' visual attention, cognitive judgments, and contextual cues into ordered visual cognitive subgraphs. Then, it unifies cognitive concepts with similar meanings but different expressions or structural roles among different experts, extracts stable consensus cognitive paths from the unified concept space, and uses these consensus cognitive paths as a reference to determine expert-specific divergent paths and their boundaries. Finally, it constructs a welded visual implicit knowledge graph that simultaneously expresses consensus cognition, divergent cognition, and visual contextual relationships. First, it unifies multi-expert cognitive concepts, solving the problem of difficulty in directly comparing cognitive nodes of different experts. Then, it extracts consensus cognitive paths, solving the problem of difficulty in stably determining the common cognitive process when multiple cognitive chains have backtracking, local reordering, and path jumps. Finally, it extracts divergent paths and constructs a unified graph, solving the problem of unclear divergent path boundaries and their structural affiliation with consensus cognitive paths. Attached Figure Description

[0093] Figure 1 A flowchart illustrating the consensus divergence path alignment welding visual implicit knowledge graph construction method provided in this embodiment of the invention;

[0094] Figure 2 Another flowchart illustrating the consensus divergence path alignment welding visual implicit knowledge graph construction method provided in this embodiment of the invention;

[0095] Figure 3 A schematic diagram illustrating the process of unifying multi-expert candidate cognitive concepts for relationship awareness, provided in an embodiment of the present invention;

[0096] Figure 4 A schematic diagram of the process for extracting multi-expert consensus cognitive paths with sequential constraints provided in an embodiment of the present invention;

[0097] Figure 5 A schematic diagram illustrating the process of consensus anchor point constraint divergence path extraction and unified graph construction provided in an embodiment of the present invention;

[0098] Figure 6 A schematic diagram of the structure of the consensus divergence and unification map of welding visual implicit knowledge provided in an embodiment of the present invention;

[0099] Figure 7 This is a schematic diagram of the fusion of visual implicit knowledge graphs for multi-object welding provided in an embodiment of the present invention. Detailed Implementation

[0100] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the embodiments of the present invention and the specific features in the embodiments are detailed descriptions of the technical solution of the present invention, rather than limitations thereof. In the absence of conflict, the embodiments of the present invention and the technical features in the embodiments can be combined with each other.

[0101] The term "and / or" simply describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A alone, A and B simultaneously, or B alone. Additionally, the character " / " generally indicates that the preceding and following related objects have an "or" relationship.

[0102] Example 1

[0103] like Figure 2 As shown in the figure, this embodiment introduces a method for constructing a welding visual implicit knowledge graph with consensus divergence path alignment, including:

[0104] Obtain visual cognitive records from several experts regarding the same welding visual inspection object, and construct a directed cognitive subgraph;

[0105] For nodes in a directed cognitive subgraph, an initial node representation is established. Then, relation-aware information is propagated across the relations in the directed cognitive subgraph to obtain relation-aware node representations.

[0106] The matching cost is calculated for the cognitive nodes represented by the relation-aware nodes. The cross-expert node correspondence is determined by one-to-one matching. The cross-expert node correspondence is checked for multi-expert consistency to obtain candidate cognitive concepts and the mapping relationship from cognitive nodes to candidate cognitive concepts.

[0107] The expert coverage is calculated based on the occurrence of candidate cognitive concepts in each expert cognitive chain. Consensus candidate concepts are determined based on the expert coverage and a preset consensus threshold. Based on the mapping relationship between cognitive nodes and candidate cognitive concepts, each expert cognitive chain is projected into the candidate cognitive concept space.

[0108] In the common expert cognitive chain, the order relationship between consensus candidate concepts is compared. Based on the order consistency rate and the preset order threshold, the cross-expert stable order relationship is determined, an order directed graph is established, and continuous directed paths that satisfy the expert coverage constraint and cognitive order constraint in the order directed graph are extracted to obtain the consensus cognitive path.

[0109] Consensus anchors corresponding to consensus cognitive paths in the expert cognitive chain are determined by sequential consistency matching. Continuous cognitive segments that are not absorbed by the consensus cognitive path between adjacent consensus anchors are extracted to form divergent paths and their left and right boundaries. Based on the consensus cognitive path, divergent paths and their left and right boundaries, and expert visual context relationships, a welding visual implicit knowledge graph is constructed.

[0110] This embodiment addresses the situation where the same cognitive concept may have differences in expression, connection relationship, and cognitive role in different expert cognitive chains. By comprehensively considering the semantic information and relational role of cognitive nodes, it performs unified matching on candidate cognitive concepts from different sources and constrains the consistency of correspondence between multiple expert sources to reduce conflicts caused by local pairwise matching, improve the accuracy of unification of multi-expert visual cognitive nodes and determination of relationships, and establish a consistent conceptual foundation for subsequent graph construction.

[0111] To address the sequence disturbances during welding visual inspection, such as back-viewing, local sequence changes, and path jumps, this study comprehensively considers the source coverage of different expert cognitive chains, the relative order between cognitive nodes, and the continuity of paths. It compares and determines the common cognitive processes in multiple cognitive chains to avoid consensus misjudgments or omissions caused by relying solely on node co-occurrence or local sequence overlap. This results in a continuous consensus cognitive path that reflects the common cognitive processes of multiple experts.

[0112] To address the situation where different experts make differing judgments due to specific visual cues during the process of shared cognition, this paper uses the established consensus cognition position as a structural reference to identify the different cognitive paths formed by each expert after that position. It also combines the preceding and following cognitive relationships to determine the starting and ending ranges of the divergent paths and their corresponding consensus positions. This allows for a clear expression of the relationship between consensus cognition paths and divergent paths within a unified graph structure, thus completing the structured organization of consensus cognition and divergent cognition in welding visual tacit knowledge.

[0113] Example 2

[0114] Based on Example 1, such as Figure 1 As shown in the figure, this embodiment introduces a method for constructing a welding visual implicit knowledge graph for consensus divergence path alignment, including the following steps:

[0115] Step 1: Unifying multi-expert candidate concepts for relationship awareness, specifically:

[0116] like Figure 3 As shown, visual cognitive records formed by several experts on the same welding visual inspection object are obtained, and a directed cognitive subgraph is constructed.

[0117] Select a welding visual inspection object to be processed. It can be a complete weld image, a region of interest in a weld, or a set of weld images that are comparable to the inspection task and visual conditions.

[0118] Obtain at least two experts' opinions on the weld visual inspection objects. Effective visual cognition records are required, with each expert's record including at least the visual attention area, the order of attention, the corresponding cognitive or judgment content, and the visual cues that triggered the cognition. When collecting eye-tracking data, the basic visual cognition sequence of the experts can be reconstructed based on fixation points, fixation duration, saccades, and area annotations.

[0119] experts Visual cognitive records are constructed into corresponding directed cognitive subgraphs. A directed cognitive subgraph contains nodes that include cognitive nodes and contextual nodes, and relationships that include at least cognitive progression relationships and context-driven relationships. To avoid symbol redundancy, directed cognitive subgraphs... The connections in the code are now uniformly called relations instead of edge set notation. .expert Corresponding directed cognitive subgraph Represented as:

[0120] ;

[0121] in, Experts The set of nodes; Experts The set of relationships. Nodes are divided into cognitive nodes and contextual nodes: cognitive nodes describe the cognitive state regarding weld morphology, defect characteristics, location features, or judgment conclusions; contextual nodes describe visual cues or detection contexts that attract expert attention. Relationships include at least two categories: cognitive-driven relationships and context-driven relationships. Cognitive nodes are organized into a sequence according to their original chronological order. And retain its original relationship direction. They represent experts respectively. The first cognitive node, the second cognitive node, the... A cognitive node; Experts The number of cognitive nodes.

[0122] Text encoding is performed on the names and descriptions of nodes in the directed cognitive subgraph to obtain semantic vectors. Simultaneously extract node types Normalized in-degree Normalized output degree Topological relative positions of nodes in the subgraph And the relative order of nodes in the expert cognitive chain For each node in the directed cognitive subgraph, an initial node representation is established. The above information is then concatenated to form the initial node representation. Initial representation of nodes Represented as:

[0123] ;

[0124] in, Indicates a node The semantic vector is obtained by text encoding the name and description; Represents a node The node type; Represents a node Normalized in-degree; Represents a node Normalized degree; Represents a node Topological relative positions in a directed cognitive subgraph; Represents a node The relative order within the expert cognitive chain; This represents vector concatenation. Context nodes do not participate in the cognitive progression order, so they can be set to... By using fixed placeholder values, this approach allows the same node to simultaneously contain textual meaning, node type, local structure, and sequential position, providing a unified input for cross-expert comparisons.

[0125] Using only textual semantics can easily merge nodes with similar names but different cognitive functions, and may also separate nodes with different expressions but at the same cognitive stage. Therefore, relation-aware information propagation is performed on the initial node representations in the directed cognitive subgraph, that is, relation-aware information propagation is performed on cognitive advancement relationships, context-driven relationships, and their reverse relationships to obtain relation-aware node representations. Let... For a set of relation types, For nodes In relation The neighbor set of nodes In the Layered relation-aware node representation Represented as:

[0126] ;

[0127] , ;

[0128] in, Represents a node In the Layer-aware node representation; Indicates the first The state transformation matrix of the nodes in the layer; Representing relations In the Layered nodes Aggregated neighborhood information; Represents a nonlinear function; Represents a node Relation-aware node representation at level 0; Indicates the first Hierarchical Relationship The transformation matrix; Represents a node In the Layer-aware node representation; This indicates taking the absolute value.

[0129] The final node is obtained after several layers of propagation. Relationship-aware node representation A two-layer propagation approach is preferred, enabling nodes to obtain direct predecessors, successors, contextual cues, and short-range contextual information, while avoiding over-propagation in smaller cognitive subgraphs.

[0130] For two different experts Take their respective cognitive nodes, and denote the semantic difference between the two cognitive nodes as . The order difference is denoted as Local structural differences are denoted as ,in, This represents the set of experts. In actual calculations, it's possible to first... , and Normalization is performed to reduce the impact of differences in the size of subgraphs perceived by different experts.

[0131] For cognitive nodes represented by relation-aware nodes, calculate the matching cost, experts Corresponding cognitive nodes With experts Corresponding cognitive nodes Matching cost between Represented as:

[0132] ;

[0133] in, They represent experts respectively. Corresponding cognitive nodes With experts Corresponding cognitive nodes Semantic differences, sequence differences, and local structural differences between them; All represent non-negative weights, satisfying ; They represent experts respectively. Corresponding cognitive nodes semantic representation vectors, experts Corresponding cognitive nodes The semantic representation vector; They represent experts respectively. Corresponding cognitive nodes The order position value, expert Corresponding cognitive nodes The sequential position value; They represent experts respectively. Corresponding cognitive nodes Local structure representation vector, expert Corresponding cognitive nodes The local structure representation vector; This represents the L2 norm.

[0134] In experts and experts One-to-one minimum-cost matching is performed between cognitive nodes to determine cross-expert node correspondences. Each node matches at most one node from another expert. The Hungarian algorithm or other equivalent minimum-cost bisection matching algorithms can be used to solve this problem. The one-to-one constraint prevents multiple nodes from being incorrectly mapped to the same node from another expert. Corresponding cognitive nodes and experts Cross-expert node correspondence between corresponding cognitive nodes Represented as:

[0135] ;

[0136] in, Experts Corresponding cognitive nodes With experts Corresponding cognitive nodes Matching indicator variables between them; This represents the candidate matching matrix corresponding to the cross-expert node correspondence that satisfies the preset matching constraints. In the middle, determine the total matching cost. The smallest matching matrix.

[0137] After completing pairwise matching of all expert source pairs, the matching results are extended to a multi-expert scope for consistency checks. Multi-expert consistency checks are performed on the cross-expert node correspondences to obtain candidate cognitive concepts and the mapping relationship from cognitive nodes to candidate cognitive concepts:

[0138] If experts Corresponding cognitive nodes and experts If there is a cross-expert node correspondence between corresponding cognitive nodes, and the two form a consistent cross-source correspondence through the node corresponding to another expert, then the cross-expert node correspondence is retained to obtain the candidate cognitive concept and the mapping relationship from cognitive node to candidate concept.

[0139] If multiple cross-expert node correspondences exist, the cross-expert node correspondence with the lowest matching cost and the highest consistency is retained to obtain candidate cognitive concepts and the mapping relationship from cognitive nodes to candidate concepts.

[0140] Cross-expert nodes that pass the consistency check are grouped into candidate cognitive concepts, forming a set of candidate cognitive concepts. Represented as:

[0141] ;

[0142] in, They represent the 1st, 2nd, and 3rd respectively. One candidate cognitive concept.

[0143] Cognitive Nodes Mapping relationship to candidate cognitive concepts Represented as:

[0144] ;

[0145] in, Indicates the first One candidate cognitive concept.

[0146] Each expert contributes at most one cognitive node within the same candidate concept. At this point, the original cognitive chains of different experts are transformed into the same candidate concept space.

[0147] Step 2: Extraction of multi-expert consensus cognitive paths with sequential constraints, specifically:

[0148] like Figure 4 As shown, expert coverage is calculated based on the occurrence of candidate cognitive concepts in each expert's cognitive chain. For each candidate cognitive concept... Statistics include the set of experts who hold this concept. and define its expert coverage, the first Expert coverage corresponding to each candidate cognitive concept Represented as:

[0149] ;

[0150] in, Indicates containing the first A set of experts for each candidate cognitive concept.

[0151] Consensus candidate concepts are determined based on expert coverage and a preset consensus threshold, forming a consensus candidate concept set. Represented as:

[0152] ;

[0153] in, This indicates a preset consensus threshold; in the preferred case, candidate concepts that appear in all experts are considered as consensus candidates, i.e., the coverage threshold is used. When there are a few missing records in the data, settings can also be configured based on the number of experts and data quality. .

[0154] This step only limits the search range of nodes for consensus cognition paths and does not directly regard nodes that appear together as consensus cognition paths. The relative order in each expert cognition chain still needs to be checked later.

[0155] Based on the mapping relationship between cognitive nodes and candidate cognitive concepts, each expert's cognitive chain is projected onto the candidate cognitive concept space, and then... experts Cognitive node sequence Projected onto the candidate cognitive concept space, experts projected onto the candidate cognitive concept space. cognitive chain Represented as:

[0156] ;

[0157] in, Experts The first in the cognitive chain A cognitive node; Experts The number of cognitive nodes in the cognitive chain, i.e., experts The number of cognitive nodes; Experts The first in the cognitive chain The mapping relationship between cognitive nodes and candidate cognitive concepts.

[0158] For candidate cognitive concepts Record it in the experts The first occurrence position in the cognitive chain is The use of relative order rather than position-by-position forced alignment allows for additional observation steps, retrospectives, or local chain length differences among different experts.

[0159] In a shared expert cognitive chain, the order of consensus candidate concepts is compared. Based on the order consistency rate and a preset order threshold, a stable cross-expert order relationship is determined, and a directed order graph is constructed. For any two consensus candidate concepts... We only compare the order of precedence among experts who appear in both cases. Let the set of common experts be denoted as . ,definition Prior to If the order consistency rate is , then the ... The consensus candidate concept and the first Order consistency rate among consensus candidate concepts Represented as:

[0160] ;

[0161] in, Indicates containing the first The consensus candidate concept and the first A common set of experts for a consensus candidate concept; They represent the first Consensus candidate concepts , No. Consensus candidate concepts In experts The position where it first appears in the cognitive chain; Indicates an indicator function.

[0162] Cross-expert stable order relations Represented as:

[0163] ;

[0164] in, Indicates the preset order threshold; preferred selection That is, a system is established only when all experts share the same direction and priorities. point to The stable order relationship; when a small amount of order perturbation is allowed, the order can be appropriately reduced. .

[0165] This allows us to construct a sequential directed graph. Concepts that do not meet the stable order condition are not subject to mandatory relationships, thus avoiding the direct writing of local reordering into the consensus backbone.

[0166] Extracting continuous directed paths from the sequential directed graph that satisfy expert coverage constraints and cognitive order constraints yields the consensus cognitive paths; the consensus backbone is defined as the sequential directed graph. The longest directed path satisfying the stable order constraint in the ordered directed graph is calculated using dynamic programming, with the path length ending at a consensus candidate concept. Consensus candidate concepts Path length to the destination Represented as:

[0167] ;

[0168] in, Indicates the first Consensus candidate concepts The path length to the destination; Indicates the first Consensus candidate concepts The set of predecessors, if Without a precursor, then... ; This indicates taking the maximum value.

[0169] Select the consensus candidate concept corresponding to the longest path length and perform reverse backtracking to obtain the consensus cognition path. Represented as:

[0170] ;

[0171] in, These respectively represent the first consensus candidate concept, the second consensus candidate concept, and the third consensus candidate concept in the consensus cognition path. One consensus candidate concept; This indicates the length of the consensus cognition path.

[0172] The consensus cognition path has a clear starting point, a continuous cognitive advancement relationship, and an end point, and is used to represent the stable common cognition process of multiple experts on the same welding visual inspection object.

[0173] Step 3: Extraction of divergent paths and construction of a unified graph based on consensus anchor constraints, specifically:

[0174] like Figure 5 As shown, after obtaining the consensus cognitive path among multiple experts, this consensus cognitive path is used as a unified reference for the visual cognitive chains of each expert. Consensus anchor points in the expert cognitive chains that maintain the same order as the consensus cognitive path are identified, and continuous cognitive segments between adjacent consensus anchor points that are not absorbed by the consensus cognitive path are extracted as divergent paths. Based on this, the consensus cognition and divergent cognition are organized into a unified welding visual implicit knowledge graph by combining the generation location, internal cognitive order, return location, and corresponding visual context of the divergent paths.

[0175] Let the consensus cognition path be ,in, The first step in the consensus cognition path A consensus candidate concept , This represents the total number of consensus candidate concepts, which is also the length of the consensus cognition path. (Expert) The cognitive chain is .

[0176] Because different experts may have missed steps in the visual inspection of welding due to retrospective visits, additional observations, or missing parts of their understanding, a position-by-position forced matching method is not used. Instead, sequential consistency matching is used to determine the consensus anchor point in the expert's cognitive chain that corresponds to the consensus cognitive path.

[0177] Calculate the maximum order matching length between candidate cognitive concepts in the expert cognitive chain and consensus candidate concepts in the consensus cognitive path. In the cognitive chain Candidate cognitive concepts mapped from each cognitive node and the consensus cognitive path (previous) Maximum order matching length among consensus candidate concepts Represented as:

[0178] ;

[0179] ;

[0180] in, Experts In the cognitive chain Candidate cognitive concepts mapped from each cognitive node and the consensus cognitive path (previous) The maximum order matching length among consensus candidate concepts; Experts In the cognitive chain Candidate cognitive concepts mapped from each cognitive node and the consensus cognitive path (previous) The maximum order matching length among consensus candidate concepts; Experts In the cognitive chain Candidate cognitive concepts mapped from each cognitive node and the consensus cognitive path (previous) The maximum order matching length among consensus candidate concepts.

[0181] Backtracking is performed based on the maximum order matching length to determine the consensus anchor points in the expert cognitive chain corresponding to the consensus cognitive path. Each consensus anchor point records both its position in the expert cognitive chain and its position in the consensus cognitive path, maintaining a consistent relative order between the two. The consensus anchor points obtained in this way constitute the structural correspondence between the expert cognitive chain and the consensus cognitive path, and serve as a reference for determining the boundaries of subsequent divergent paths.

[0182] By sequentially examining the cognitive nodes between adjacent consensus anchors in the expert cognitive chain, and extracting continuous cognitive fragments between adjacent consensus anchors that were not absorbed by the consensus cognitive path, a divergence path and its left and right boundaries are formed. The corresponding number Divergent paths Represented as:

[0183] ;

[0184] in, They represent experts respectively. The corresponding number The starting and ending positions of each diverging path in the cognitive chain.

[0185] Further define the left and right boundaries of the divergence path. The left boundary of the divergence path is the consensus anchor point location corresponding to the left side of the divergence path. The right boundary of the divergence path is the location of the consensus anchor point corresponding to the right side of the divergence path. .

[0186] A divergence path located between two consensus anchors has left and right boundaries. A divergence path occurring before the first cognitive node in a consensus cognitive path has a right boundary. A divergence path occurring after the last cognitive node in a consensus cognitive path and not re-entering the subsequent consensus cognitive process has a left boundary.

[0187] Through the above processing, differential cognition can be transformed from discrete nodes into a continuous cognitive path with a clear starting position, internal order, and ending position.

[0188] Based on consensus-based cognitive paths, divergent paths and their left and right boundaries, and expert visual context relationships, a welding visual implicit knowledge graph is constructed:

[0189] When a divergence path occurs after a cognitive node in the consensus cognitive path, a divergence trigger edge is formed by pointing from the cognitive node in that consensus cognitive path to the cognitive node at the beginning of the divergence path. .

[0190] The divergent paths are formed within each path according to the order of expert cognition, in order to maintain the continuity of the expert's differential cognitive process.

[0191] When a divergent path re-enters the subsequent consensus cognition process, a divergent return edge is formed, pointing from the cognition node at the end of the divergent path to the cognition node in the subsequent consensus cognition path. .

[0192] For divergent paths with consensus anchors on the left and right, their structural relationship can be represented as:

[0193] .

[0194] in, The first step in the consensus cognition path One consensus candidate concept, namely the left boundary of the divergence path. Consensus candidate concepts; The first step in the consensus cognition path One consensus candidate concept, namely the right boundary of the divergence path. Consensus candidate concepts.

[0195] For divergent paths with only a one-sided consensus anchor, only connections corresponding to the actual cognitive structure are established, without attaching non-existent consensus relationships. This maintains the structural affiliation between divergent cognition and its corresponding stage of consensus cognition.

[0196] After determining the consensus and divergence paths, a visual implicit knowledge graph for welding is constructed. The cognitive nodes in the consensus cognitive path are connected through cognitive advancement relationships according to the cognitive order of the consensus cognitive path. The divergent cognitive path is generated from the cognitive nodes in the consensus cognitive path through divergence triggering edges, and the cognitive order is maintained through divergence path edges. When re-entering the consensus cognitive process, it is connected to the cognitive nodes in the subsequent consensus cognitive path through divergence return edges.

[0197] By preserving the context nodes and context-driven relationships in the directed cognitive subgraph, a well-constructed visual implicit knowledge graph of welding is obtained. .

[0198] Welding visual implicit knowledge graph The nodes include consensus cognition nodes, divergence cognition nodes, consensus context nodes, and divergence context nodes.

[0199] Consensus cognition nodes are used to represent a stable cognitive process jointly followed by multiple experts, divergence cognition nodes are used to represent a differential cognitive process formed by a specific expert at a corresponding consensus position, consensus context nodes are used to represent the weld-like visual cues that support consensus cognition, and divergence context nodes are used to represent the visual cues that trigger or support divergent cognition among experts.

[0200] Directed relations in the graph include consensus-progressing edges. Divergence trigger edge Branching path edge, branching return edge Consensus Context Edge and the edge of the divergent situation .

[0201] Consensus nodes pass through the consensus path in the order of their cognition. Connection, divergent cognitive paths through It originates from the corresponding consensus position and maintains the original cognitive order through the divergence path edges. Upon re-entering the consensus cognitive process, it... Connects to subsequent consensus-building nodes. Consensus advancement edge. This is a cognitive advancement relationship.

[0202] Context nodes and context-driven relationships in the directed cognitive subgraph are preserved synchronously. Visual contexts supporting consensus cognition are... Connecting with corresponding consensus-based cognitive nodes, the visual context supporting the differentiated perceptions of experts is achieved through... Connect with corresponding divergent cognitive nodes. The same visual context can be associated with one or more cognitive nodes to maintain the correspondence between the original visual cues and the cognitive judgments.

[0203] The final weld visual implicit knowledge graph A stable backbone is formed by consensus-based cognitive paths, and branch structures with clear boundaries are formed by expert-disagreement-based cognitive paths. The corresponding visual context and expert source information are retained, thereby forming a welding visual implicit knowledge graph that unifies the relationship between consensus-based cognitive structures, disagreement-based cognitive structures and visual context.

[0204] like Figure 6 As shown, consensus cognition nodes form a stable consensus cognition backbone along the horizontal direction, representing the welding visual inspection process shared by multiple experts; consensus cognition nodes are connected through consensus advancement relationships. The consensus context nodes corresponding to the consensus cognition nodes are used to record the welding visual cues supporting the consensus cognition, such as the weld start position, weld toe area, weld surface, and weld quality judgment.

[0205] Disagreement nodes are located on either side of the consensus backbone and are generated from the corresponding consensus nodes through disagreement triggering relationships. For expert differing cognitive processes involving multiple consecutive cognitive steps, adjacent disagreement nodes are connected through disagreement path relationships; when a disagreement re-enters the consensus process, it is connected to subsequent consensus nodes through disagreement return relationships.

[0206] The divergent context nodes are used to record the specific visual contexts that trigger or support divergent cognitions, and are connected to the corresponding divergent cognition nodes through divergent context relationships. Therefore, this diagram can simultaneously express stable cognitive paths commonly followed by multiple experts, expert-specific and differentiated cognitive paths, and the corresponding visual context relationships.

[0207] like Figure 7 As shown, the fusion results of object-level knowledge graphs formed by multiple welding visual inspection objects in a unified graph space are presented, such as... Figure 6 As shown, the mapping of a specific object in the global graph is given.

[0208] Figure 7 This represents the distribution of the constructed results of multiple objects in a unified graph space. Different node types represent consensus cognition, divergent cognition, consensus context, and divergent context, respectively. The highlighted node represents the corresponding node of a specific welding inspection object in the global graph. Through the mapping relationship between object-level nodes and global cognitive concepts, the position of the object's cognitive structure in the overall knowledge space can be observed.

[0209] Figure 6This represents the object-level graph of the selected object after global concept mapping. Consensus cognitive nodes form a continuous backbone relationship, while divergent cognitive nodes are connected to the consensus backbone through divergence triggering, divergence advancement, and divergence termination or return relationships. Various contextual nodes are connected to the corresponding cognitive nodes through contextual relationships.

[0210] Example 3

[0211] Based on Example 2, this example presents an experimental example of a consensus-based divergence path alignment method for constructing a visual implicit knowledge graph for welding:

[0212] The same weld image or a local area of ​​the same weld was selected as the inspection object, and multiple welding inspection experts conducted independent inspections. The visual focus areas, order of focus, cognitive or judgment content, and visual cues that triggered the corresponding cognition were collected for each expert during the inspection process, and the experts' visual cognitive process was reconstructed in chronological order.

[0213] Each expert's visual cognitive record is converted into a directed cognitive subgraph. The cognitive subgraph includes cognitive nodes and context nodes. Cognitive nodes represent the expert's cognitive state regarding the weld location, morphology, defect characteristics, and judgment results. Context nodes represent the visual cues that trigger or support the corresponding cognition. Cognitive progression relationships are established between adjacent cognitive nodes, and context-driven relationships are established between context nodes and the cognitive nodes they support.

[0214] This embodiment selects five expert cognitive subgraphs corresponding to the same inspection object for verification. The five cognitive subgraphs contain a total of 54 nodes and 65 directed relationships, including 35 cognitive nodes, 19 situational nodes, 30 cognitive progression relationships, and 35 situational driving relationships. The cognitive subgraphs formed by different experts differ in the number of nodes and the length of cognitive paths, which can reflect the additional observations, local judgments, and path differences of experts in actual welding visual inspection.

[0215] First, an initial representation of each node in the cognitive subgraph is established. Node names and descriptions are semantically encoded, and node type, in-degree, out-degree, topological position, and relative order of the cognitive node within the expert cognitive chain are extracted. This information is then combined to form a unified initial node representation. Contextual nodes do not participate in the cognitive chain order but are retained in the cognitive subgraph and participate in the propagation of subsequent relationship information through context-driven relationships.

[0216] In this embodiment, the initial node representation consists of a 384-dimensional text semantic vector, a 2-dimensional node type encoding, and four structural quantities: in-degree, out-degree, topological position, and cognitive order, forming a total of 390-dimensional initial node representation. Subsequently, a two-layer relation-aware graph neural network is used to propagate information across cognitive progression relationships, context-driven relationships, and their reverse relationships, resulting in a 128-dimensional relation-aware node representation, enabling nodes to simultaneously contain semantic information, cognitive order information, and local graph structure information.

[0217] For cognitive nodes from different experts, a comprehensive matching cost is calculated based on differences in node semantics, cognitive order, and local structure. A one-to-one minimum cost matching method is then used to determine the correspondence between candidate nodes from different experts. This one-to-one constraint ensures that a cognitive node from one expert corresponds to at most one cognitive node from another expert, preventing multiple nodes from being incorrectly merged into the same cognitive concept.

[0218] After pairwise matching of experts, a consistency check is performed on the cross-expert node correspondences. When the matching between two nodes can be kept consistent through corresponding nodes in other experts, the correspondence is retained. For conflicting matching results, the relationship with the lower matching cost and the ability to maintain multi-expert consistency is retained. Finally, the cognitive nodes that pass the consistency check are merged into unified candidate cognitive concepts, and the correspondence between the original expert cognitive nodes and the candidate cognitive concepts is recorded.

[0219] In this embodiment, after relation-aware representation, one-to-one matching and multi-expert consistency checks, the 35 cognitive nodes are unified into 12 candidate cognitive concepts and form 56 candidate matching relationships.

[0220] After unifying the candidate cognitive concepts, the occurrence of each candidate cognitive concept in different expert cognitive chains is statistically analyzed, and the corresponding expert coverage is calculated. Candidate cognitive concepts appearing in all expert chains are preferred as consensus candidate concepts; for cases with a small number of missing records, a corresponding consensus threshold can be set based on the number of valid experts.

[0221] Based on the mapping relationship between the original cognitive nodes and candidate cognitive concepts, each expert's cognitive chain is projected onto a unified candidate concept space. This process only unifies the conceptual representation of the nodes without changing the order of the original cognitive chains of the experts, thus preserving the additional observations, retrospectives, and local insertions that occur during the actual inspection process.

[0222] Compare the relative order of any two consensus candidate concepts in different expert cognitive chains. Establish a stable order relationship when the two concepts maintain a stable order direction in their co-occurring expert cognitive chains; for candidate concepts with obvious order conflicts, do not establish a forced order relationship. Construct a directed graph based on the stable order relationships.

[0223] In a sequential directed graph, the longest continuous directed path that satisfies expert coverage and stable order constraints is searched using dynamic programming, and the multi-expert consensus cognitive path is determined through backtracking. In this embodiment, five consensus cognitive nodes with full expert source coverage and stable order are extracted from 12 candidate cognitive concepts, forming a continuous consensus cognitive backbone.

[0224] After obtaining the consensus cognitive path, the consensus path is used as a unified reference for each expert's cognitive chain. The candidate concept sequence of each expert is matched sequentially to determine the cognitive nodes in the expert cognitive chain that can maintain relative sequential consistency with the consensus path, and the corresponding positions are recorded as consensus anchor points.

[0225] The cognitive nodes between adjacent consensus anchors are examined sequentially. When there are consecutive cognitive nodes between two consensus anchors that have not been absorbed by the consensus path, this consecutive cognitive segment is identified as a divergence path. Each divergence path records its expert source, internal cognitive order, and the positions of its left and right consensus anchors, thereby clarifying which consensus cognitive stage generated the divergent cognition and where it re-enters the consensus process.

[0226] For divergent paths with left and right consensus anchors, a divergence triggering relationship is established by pointing the left consensus cognitive node to the first cognitive node of the divergent path; within the divergent path, divergent path relationships are established according to the original cognitive order of experts; when the end of the divergent path re-enters the consensus process, a divergence return relationship is established by pointing the end divergent cognitive node to the right consensus cognitive node.

[0227] In this embodiment, the 10 cognitive nodes outside the consensus backbone are organized into 7 branching paths, of which 5 branching paths contain 1 cognitive node, 1 contains 2 cognitive nodes, and 1 contains 3 cognitive nodes. All 7 branching paths have clear left and right consensus anchors, can return to subsequent consensus nodes, and no sequential conflict of cognitive concepts in the consensus backbone is detected.

[0228] After the consensus path and the divergence path are determined, they are organized in a unified manner with the visual context relationship in the original expert cognitive subgraph to construct a welding visual implicit knowledge graph.

[0229] The nodes in the graph include consensus cognitive nodes, divergent cognitive nodes, consensus context nodes, and divergent context nodes. Among them, consensus cognitive nodes represent a stable cognitive process formed jointly by multiple experts, divergent cognitive nodes represent a differentiated cognitive process formed by a specific expert, consensus context nodes represent visual cues that support consensus cognitive processes, and divergent context nodes represent visual cues that trigger or support divergent cognitive processes among experts.

[0230] The directed relationships in the graph include consensus advancement relationships, divergence triggering relationships, divergence path relationships, divergence return relationships, consensus context relationships, and divergence context relationships. Consensus cognitive nodes are connected according to the actual cognitive order of the consensus path to form the consensus backbone; divergence paths are generated from the corresponding consensus nodes and reconnected to subsequent consensus nodes when they have a right-hand consensus anchor point; visual context nodes are connected to the corresponding cognitive nodes according to the context-driven relationships in the original expert cognitive subgraph.

[0231] A single visual context can simultaneously support one or more consecutive cognitive nodes. Therefore, the original one-to-many contextual relationships are preserved during the graph construction process, and contextual nodes are not forcibly inserted into the cognitive progression path. Through this approach, consensus-based cognitive processes, expert-specific cognitive processes, and corresponding visual contexts have clear structural boundaries and relationships within the unified graph structure.

[0232] The unified graph obtained in this embodiment includes a continuous 5-node consensus cognitive backbone, 10 divergent cognitive nodes and corresponding context nodes, forming 4 consensus advancement relationships, 7 divergence triggering relationships, 3 divergence path relationships, 7 divergence return relationships and 15 context association relationships.

[0233] To verify the effectiveness of the relation-aware node representation, typical cognitive nodes were compared using both plain text semantic representation and relation-aware representation. For nodes with the same text content but different structural roles in the cognitive subgraph, the distance under plain text representation was 0. After adding cognitive relation and local structural information, the distance increased to 0.132, indicating that relation-aware representation can distinguish nodes with the same text but different cognitive functions. For nodes with different textual expressions but similar cognitive order and local relational roles, the distance decreased from 0.618 to 0.001, indicating that the method can identify nodes with similar cognitive functions under different linguistic expressions.

[0234] To verify the stability of the relation-aware encoding process, the experimental data was divided into training, validation, and test sets according to the welding inspection objects, avoiding the simultaneous appearance of different expert cognitive subgraphs of the same object in different datasets. The experimental corpus contained 750 cognitive subgraphs, 9375 nodes, and 11790 relations, of which 595 cognitive subgraphs were used for training, 75 for validation, and 80 for testing.

[0235] After 35 training cycles, the separability of relations on the training, validation, and test sets reached 0.881, 0.864, and 0.874, respectively; the average predicted probability of true relations on the test set was 0.922, and the average predicted probability of negative relations was 0.049. These results demonstrate that the adopted relation-aware node representation can effectively utilize node semantics, cognitive order, and local relation structure to form a stable node relation representation.

[0236] From the overall mapping results, this embodiment sequentially completed the unification of 35 cognitive nodes into 12 candidate cognitive concepts, the extraction of 5 stable consensus cognitive nodes, and the organization of 10 remaining cognitive nodes into 7 boundary-bounded divergent paths. Each processing stage preserved the original expert sources and node correspondences, ensuring that the consensus cognitive nodes, divergent cognitive nodes, and visual contextual relationships in the final graph could be mapped to the original expert cognitive records.

[0237] The above-mentioned atlas construction process was further performed on 150 welding visual inspection objects, resulting in a total of 1,642 candidate cognitive concepts, 882 consensus backbone nodes, 1,389 divergent paths, and 1,860 divergent cognitive nodes. This shows that the present invention can continuously complete the unification of candidate concepts, extraction of consensus paths, and structured construction of divergent paths under multiple inspection object conditions.

[0238] In summary, this invention can form a unified candidate cognitive concept, extract a consensus cognitive path with a stable sequential relationship, and organize expert-specific cognition into divergent paths with clear start and end boundaries, even when there are differences in textual expression, cognitive chain length, and local sequence changes in the visual cognitive records of multiple welding inspection experts. This allows for the construction of a welding visual implicit knowledge graph that unifies the expression of consensus cognition, divergent cognition, and visual contextual relationships.

[0239] Example 4

[0240] Based on Example 2, this example presents an experimental example of a consensus-based divergence path alignment method for constructing a visual implicit knowledge graph for welding:

[0241] This embodiment unifies multi-expert candidate cognitive concepts by combining node semantics, cognitive order, and local relational structure. This can reduce the mis-merging and mis-separation of concepts caused by matching based solely on text semantics or local neighborhoods, and improve the consistency of cross-expert cognitive node correspondence.

[0242] To verify the effectiveness of this method, we set up comparison methods 1 (graph convolution-based node representation), 2 (neighborhood sampling aggregation-based graph representation), and 3 (graph attention-based node representation) for comparison. The results of the comparison experiments are shown in Table 1.

[0243] Table 1 Comparative test results

[0244]

[0245] Experimental results show that our proposed method achieves ROC-AUC, AP, F1, MRR, Hits@1, and Hits@3 scores of 0.7744, 0.7899, 0.7430, 0.7102, 0.5656, and 0.8156, respectively, outperforming the three comparative methods overall. Compared to the best-performing comparative method 3, our proposed method improves AP, MRR, and Hits@1 scores by 0.1043, 0.1413, and 0.2459, respectively, indicating that relation-aware representation can more accurately distinguish different cognitive roles and improve the stability of cross-expert candidate concept matching.

[0246] This invention extracts consensus cognitive paths by considering expert coverage relationships, stable order, and path search in a directed graph of candidate concepts. This avoids consensus misjudgments caused by relying solely on node co-occurrence or local sequence overlap, and reduces the impact of backtracking, local reordering, and additional observation steps during expert visual inspection on consensus extraction. In a specific embodiment, among the 12 candidate cognitive concepts formed after unifying the candidate concepts, 5 consensus cognitive nodes with stable expert coverage and cognitive order are further identified, forming a continuous consensus backbone. This allows the cognitive process shared by multiple experts to be expressed as a complete path rather than discrete nodes.

[0247] This invention uses consensus paths as a structural reference, identifying continuous cognitive segments in the expert cognitive chain that are not absorbed by the consensus backbone through consensus anchor points. These segments are then organized into branching paths with clearly defined start and end points, preventing the simple deletion, isolated storage, or mixing of expert discrepancies with the consensus structure. In a specific embodiment, 10 cognitive nodes outside the consensus backbone are organized into 7 branching paths. Each branching path has a clear consensus anchor point and structural affiliation, and can be connected to a unified graph through branching trigger relationships, branching path relationships, and branching return relationships. Therefore, the constructed welding visual implicit knowledge graph can simultaneously and clearly express consensus cognition, expert discrepancies in cognition, and corresponding visual context relationships, improving the integrity and interpretability of the graph structure.

[0248] After constructing the consensus divergence graph for a single welding inspection object, this invention can further construct a globally unified graph for multiple object-level objects. Specifically, by synchronizing candidate cognitive concepts and contextual concepts across different objects, object-level nodes with similar relationships, relative positions, and local topological roles are merged into a global cognitive prototype, and repetitive relationships are uniformly integrated. In the experiment, object-level graphs were constructed for 150 welding inspection objects, resulting in a total of 1642 candidate cognitive concepts, 882 consensus backbone nodes, 1389 divergence paths, and 1860 divergence cognitive nodes. After further global fusion, 1102 global cognitive concepts and 1835 global contextual concepts were obtained, ultimately forming a global welding visual implicit knowledge graph containing 2937 nodes and 4619 relationships. The results show that the present invention can not only maintain the integrity of the consensus and divergence structure within a single object, but also unify and merge repetitive cognitive expressions under multi-object conditions, and maintain the correspondence between object source, expert source and local structure, thereby verifying the scalability and structural stability of the proposed graph construction method under large-scale data conditions.

[0249] Example 5

[0250] Based on embodiments 1 to 4, this embodiment introduces a consensus-based divergence path alignment device for constructing a welding visual implicit knowledge graph, comprising:

[0251] The subgraph construction module is used to: acquire visual cognitive records formed by several experts on the same welding visual inspection object, and construct a directed cognitive subgraph;

[0252] The node representation module is used to: establish initial node representations for nodes in a directed cognitive subgraph, propagate relation-aware information on the initial node representations in the relation of the directed cognitive subgraph, and obtain relation-aware node representations;

[0253] The matching and checking module is used to: calculate the matching cost for cognitive nodes represented by relation-aware nodes, determine the cross-expert node correspondence through one-to-one matching, perform multi-expert consistency checks on the cross-expert node correspondence, and obtain candidate cognitive concepts and the mapping relationship from cognitive nodes to candidate cognitive concepts.

[0254] The concept mapping module is used to: calculate expert coverage based on the occurrence of candidate cognitive concepts in each expert cognitive chain; determine consensus candidate concepts based on expert coverage and a preset consensus threshold; and project each expert cognitive chain to the candidate cognitive concept space based on the mapping relationship from cognitive nodes to candidate cognitive concepts.

[0255] The consensus extraction module is used to: compare the order of consensus candidate concepts in the common expert cognitive chain, determine the cross-expert stable order relationship based on the order consistency rate and the preset order threshold, establish an ordered directed graph, extract the continuous directed paths in the ordered directed graph that satisfy the expert coverage constraint and the cognitive order constraint, and obtain the consensus cognitive path;

[0256] The graph construction module is used to: determine the consensus anchor points in the expert cognitive chain corresponding to the consensus cognitive path through sequential consistency matching; extract continuous cognitive segments between adjacent consensus anchor points that are not absorbed by the consensus cognitive path to form divergent paths and their left and right boundaries; and construct a welding visual implicit knowledge graph based on the consensus cognitive path, divergent paths and their left and right boundaries, and expert visual context relationships.

[0257] The specific functions of each module described above are explained in the relevant content of the method in Embodiment 1, and will not be repeated here.

[0258] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0259] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0260] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0261] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0262] The embodiments of the present invention have been described above with reference to the accompanying drawings. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of the present invention without departing from the spirit and scope of the claims. All of these forms are within the protection scope of the present invention.

Claims

1. A method for constructing a visual implicit knowledge graph for welding based on consensus divergence path alignment, characterized in that, include: Obtain visual cognitive records from several experts regarding the same welding visual inspection object, and construct a directed cognitive subgraph; For nodes in a directed cognitive subgraph, an initial node representation is established. Then, relation-aware information is propagated on the relations in the directed cognitive subgraph to obtain relation-aware node representations. The matching cost is calculated for the cognitive nodes represented by the relation-aware nodes. The cross-expert node correspondence is determined by one-to-one matching. The cross-expert node correspondence is checked for multi-expert consistency to obtain candidate cognitive concepts and the mapping relationship from cognitive nodes to candidate cognitive concepts. The expert coverage is calculated based on the occurrence of candidate cognitive concepts in each expert cognitive chain. Consensus candidate concepts are determined based on the expert coverage and a preset consensus threshold. Based on the mapping relationship between cognitive nodes and candidate cognitive concepts, each expert cognitive chain is projected into the candidate cognitive concept space. In the common expert cognitive chain, the order relationship between consensus candidate concepts is compared. Based on the order consistency rate and the preset order threshold, the cross-expert stable order relationship is determined, an order directed graph is established, and continuous directed paths that satisfy the expert coverage constraint and cognitive order constraint in the order directed graph are extracted to obtain the consensus cognitive path. Consensus anchors corresponding to consensus cognitive paths in the expert cognitive chain are determined by sequential consistency matching. Continuous cognitive segments that are not absorbed by the consensus cognitive path between adjacent consensus anchors are extracted to form divergent paths and their left and right boundaries. Based on the consensus cognitive path, divergent paths and their left and right boundaries, and expert visual context relationships, a welding visual implicit knowledge graph is constructed.

2. The method for constructing a welding visual implicit knowledge graph based on consensus divergence path alignment according to claim 1, characterized in that, The visual cognition record includes at least the visual attention area, the order of attention, the corresponding cognition or judgment content, and the visual cues that trigger the cognition. The nodes in the directed cognitive subgraph include cognitive nodes and context nodes, and the relationships include at least cognitive advancement relationships and context-driven relationships.

3. The method for constructing a welding visual implicit knowledge graph based on consensus divergence path alignment according to claim 1, characterized in that, The initial representation of the node is as follows: ； in, Represents a node The initial representation of the nodes; Indicates a node The semantic vector is obtained by text encoding the name and description; Represents a node The node type; Represents a node Normalized in-degree; Represents a node Normalized degree; Represents a node Topological relative positions in a directed cognitive subgraph; Represents a node The relative order within the expert cognitive chain; This indicates vector concatenation; The relation-aware node representation is as follows: ； ， ； in, Representing nodes respectively In the Layer, First Layer-aware node representation; Indicates the first The state transformation matrix of the nodes in the layer; Represents a set of relation types; Representing relations In the Layered nodes Aggregated neighborhood information; Represents a nonlinear function; Represents a node Relation-aware node representation at level 0; Represents a node In relationship The set of neighbors below; Indicates the first Hierarchical Relationship The transformation matrix; Represents a node In the Layer-aware node representation; This indicates taking the absolute value.

4. The method for constructing a welding visual implicit knowledge graph based on consensus divergence path alignment according to claim 1, characterized in that, The matching cost is expressed as: ； ， ， ； in, Experts Corresponding cognitive nodes With experts Corresponding cognitive nodes The matching cost between them; They represent experts respectively. Corresponding cognitive nodes With experts Corresponding cognitive nodes Semantic differences, sequence differences, and local structural differences between them; All of these represent non-negative weights; They represent experts respectively. Corresponding cognitive nodes semantic representation vectors, experts Corresponding cognitive nodes The semantic representation vector; They represent experts respectively. Corresponding cognitive nodes The order position value, expert Corresponding cognitive nodes The sequential position value; They represent experts respectively. Corresponding cognitive nodes Local structure representation vector, expert Corresponding cognitive nodes The local structure representation vector; Indicates taking the absolute value; Represents the L2 norm; The cross-expert node correspondence is represented as follows: ； in, Experts Corresponding cognitive nodes and experts The corresponding cross-expert node correspondence between cognitive nodes; Experts Corresponding cognitive nodes With experts Corresponding cognitive nodes Matching indicator variables between them; This represents the candidate matching matrix corresponding to the cross-expert node correspondence that satisfies the preset matching constraints. In the process, determine the matching matrix that minimizes the total matching cost; A multi-expert consistency check is performed on the cross-expert node correspondence to obtain candidate cognitive concepts and the mapping relationship from cognitive nodes to candidate concepts, including: If experts Corresponding cognitive nodes and experts If there is a cross-expert node correspondence between corresponding cognitive nodes, and the two form a consistent cross-source correspondence through the node corresponding to another expert, then the cross-expert node correspondence is retained to obtain the candidate cognitive concept and the mapping relationship from cognitive node to candidate concept. If multiple cross-expert node correspondences exist, the cross-expert node correspondences with the lowest matching cost and the highest consistency are retained to obtain candidate cognitive concepts and the mapping relationship from cognitive nodes to candidate concepts. The candidate cognitive concepts are represented as follows: ； in, Represents a set of candidate cognitive concepts; They represent the 1st, 2nd, and 3rd respectively. One candidate cognitive concept; The mapping relationship from the cognitive node to the candidate cognitive concept is represented as follows: ； in, Representing cognitive nodes The mapping relationship to candidate cognitive concepts; Indicates the first One candidate cognitive concept.

5. The method for constructing a welding visual implicit knowledge graph based on consensus divergence path alignment according to claim 1, characterized in that, The expert coverage is expressed as follows: ； in, Indicates the first Expert coverage corresponding to each candidate cognitive concept; Indicates containing the first A set of experts for each candidate cognitive concept; Represents a set of experts; Indicates taking the absolute value; The consensus candidate concept is represented as follows: ； in, Represents the set of consensus candidate concepts; Indicates the first One consensus candidate concept; Represents a set of candidate cognitive concepts; This indicates a preset consensus threshold; The expert cognitive chains projected onto the candidate cognitive concept space are represented as follows: ； in, Experts The cognitive chain; Experts The first in the cognitive chain A cognitive node; Experts The number of cognitive nodes in the cognitive chain; Experts The first in the cognitive chain The mapping relationship between cognitive nodes and candidate cognitive concepts.

6. The method for constructing a welding visual implicit knowledge graph based on consensus divergence path alignment according to claim 1, characterized in that, The order consistency rate is expressed as: ； in, Indicates the first The consensus candidate concept and the first The order consistency rate among consensus candidate concepts; Indicates containing the first The consensus candidate concept and the first A common set of experts for a consensus candidate concept; They represent the first Consensus candidate concepts , No. Consensus candidate concepts In experts The position where it first appears in the cognitive chain; Indicates an indicator function; Indicates taking the absolute value; The cross-expert stable order relationship is represented as follows: ； in, Indicates a stable order relationship across experts; Indicates the preset order threshold; The directed graph of the sequence is represented as: ； in, Represents a directed graph in sequence; This represents the set of consensus candidate concepts.

7. The method for constructing a welding visual implicit knowledge graph based on consensus divergence path alignment according to claim 1, characterized in that, Extracting continuous directed paths from the ordered directed graph that satisfy expert coverage constraints and cognitive order constraints yields consensus cognitive paths, including: Using dynamic programming, the path length of the consensus candidate concept in the sequential directed graph is calculated. The consensus candidate concept corresponding to the longest path length is selected and backtracked to obtain the consensus cognition path. The path length is expressed as: ； in, They respectively represent the first Consensus candidate concepts , No. Consensus candidate concepts The path length to the destination; Indicates the first Consensus candidate concepts The set of predecessors; This indicates taking the maximum value; The consensus cognition path is represented as follows: ； in, This represents a consensus-based cognitive path; These respectively represent the first consensus candidate concept, the second consensus candidate concept, and the third consensus candidate concept in the consensus cognition path. One consensus candidate concept; This indicates the length of the consensus cognition path.

8. The method for constructing a welding visual implicit knowledge graph based on consensus divergence path alignment according to claim 1, characterized in that, Consensus anchors corresponding to the consensus cognition path in the expert cognition chain are determined through sequential consistency matching, including: Calculate the maximum sequential matching length between candidate cognitive concepts in the expert cognitive chain and consensus candidate concepts in the consensus cognitive path, and perform reverse backtracking based on the maximum sequential matching length to determine the consensus anchor point in the expert cognitive chain corresponding to the consensus cognitive path. The maximum order matching length between candidate cognitive concepts in the expert cognitive chain and consensus candidate concepts in the consensus cognitive path is expressed as: ； ； in, Experts In the cognitive chain Candidate cognitive concepts mapped from each cognitive node and the consensus cognitive path (previous) The maximum order matching length among consensus candidate concepts; Experts In the cognitive chain Candidate cognitive concepts mapped from each cognitive node and the consensus cognitive path (previous) The maximum order matching length among consensus candidate concepts; Experts In the cognitive chain Candidate cognitive concepts mapped from each cognitive node and the consensus cognitive path (previous) The maximum order matching length among consensus candidate concepts; Experts In the cognitive chain Candidate cognitive concepts mapped from each cognitive node and the consensus cognitive path (previous) The maximum order matching length among consensus candidate concepts; Experts The first in the cognitive chain A cognitive node; The first step in the consensus cognition path One consensus candidate concept; This indicates taking the maximum value; The branching path is represented as follows: ； in, Experts The corresponding number A divergent path; Experts The first in the cognitive chain A cognitive node; They represent experts respectively. The corresponding number The starting and ending positions of the diverging paths in the cognitive chain; The left boundary of the divergence path is the consensus anchor point position corresponding to the left side of the divergence path, and the right boundary of the divergence path is the consensus anchor point position corresponding to the right side of the divergence path. The divergence path located between two consensus anchor points has left and right boundaries. The divergence path occurring before the first cognitive node in the consensus cognitive path has a right boundary. The divergence path occurring after the last cognitive node in the consensus cognitive path and not re-entering the subsequent consensus cognitive process has a left boundary.

9. The method for constructing a welding visual implicit knowledge graph based on consensus divergence path alignment according to claim 1, characterized in that, Based on consensus-based cognitive paths, divergent paths and their left and right boundaries, and expert visual context relationships, a welding visual implicit knowledge graph is constructed, including: When a divergent path is generated after a cognitive node in the consensus cognitive path, a divergent trigger edge is formed by pointing from the cognitive node in the consensus cognitive path to the cognitive node at the beginning of the divergent path. The divergent path edges are formed inside the divergent path according to the order of expert cognition. When a divergent path re-enters the subsequent consensus cognition process, a divergent return edge is formed by the cognitive node at the end of the divergent path pointing to the cognitive node in the subsequent consensus cognition path. The cognitive nodes in the consensus cognitive path are connected through cognitive advancement relationships according to the cognitive order of the consensus cognitive path. The divergent cognitive path is generated from the cognitive nodes in the consensus cognitive path through divergence triggering edges, and the cognitive order is maintained through divergence path edges. When re-entering the consensus cognitive process, it is connected to the cognitive nodes in the subsequent consensus cognitive path through divergence return edges. By preserving the context nodes and context-driven relationships in the directed cognitive subgraph, a well-constructed visual implicit knowledge graph of welding is obtained.

10. A device for constructing a visual implicit knowledge graph for welding based on consensus divergence path alignment, characterized in that, include: The subgraph construction module is used to: acquire visual cognitive records formed by several experts on the same welding visual inspection object, and construct a directed cognitive subgraph; The node representation module is used to: establish initial node representations for nodes in a directed cognitive subgraph, propagate relation-aware information on the initial node representations in the relation of the directed cognitive subgraph, and obtain relation-aware node representations; The matching and checking module is used to: calculate the matching cost for cognitive nodes represented by relation-aware nodes, determine the cross-expert node correspondence through one-to-one matching, perform multi-expert consistency checks on the cross-expert node correspondence, and obtain candidate cognitive concepts and the mapping relationship from cognitive nodes to candidate cognitive concepts. The concept mapping module is used to: calculate expert coverage based on the occurrence of candidate cognitive concepts in each expert cognitive chain; determine consensus candidate concepts based on expert coverage and a preset consensus threshold; and project each expert cognitive chain to the candidate cognitive concept space based on the mapping relationship from cognitive nodes to candidate cognitive concepts. The consensus extraction module is used to: compare the order of consensus candidate concepts in the common expert cognitive chain, determine the cross-expert stable order relationship based on the order consistency rate and the preset order threshold, establish an ordered directed graph, extract the continuous directed paths in the ordered directed graph that satisfy the expert coverage constraint and the cognitive order constraint, and obtain the consensus cognitive path; The graph construction module is used to: determine the consensus anchor points in the expert cognitive chain corresponding to the consensus cognitive path through sequential consistency matching; extract continuous cognitive segments between adjacent consensus anchor points that are not absorbed by the consensus cognitive path to form divergent paths and their left and right boundaries; and construct a welding visual implicit knowledge graph based on the consensus cognitive path, divergent paths and their left and right boundaries, and expert visual context relationships.

Images