Determination method for implant restoration connection compatibility based on node2vec and compatibility knowledge graph retrieval method

Abstract

The application provides a planting repair connection compatibility determination method based on Node2Vec and a compatibility knowledge graph retrieval method; the method constructs a compatibility knowledge graph containing connection form matching, platform diameter matching, screw specification matching and series intergenerational mapping by using available part main data and connection data, and defines a connection chain template in the order of the above relationships for structured retrieval; when the retrieval is interrupted, implicit similarity representation is established by template consistent random walk and representation learning by taking the breakpoint anchor and the missing relationship type as constraints, a candidate compatible relationship is generated, written into an enhanced graph and rechecked to obtain a verified connection chain, and then an explicit compatibility, an inferred compatibility or an undeterminable result is outputted, and a determination record containing the connection chain / breakpoint is formed; the method improves the determination closed loop capability and the result interpretability.

Description

Technical Field

[0001] This invention relates to the field of dental implant restoration connectivity compatibility assessment technology, and in particular to a method for assessing implant restoration connectivity compatibility based on Node2Vec and compatibility knowledge graph retrieval. Background Technology

[0002] In the assembly process of dental implant restorations, the ability to connect the implant and abutment typically depends on the consistency and mappability of interface specifications, such as connection type, platform diameter, screw specifications, and connection inheritance relationships between different generations of the same series. Clinical and supply chain scenarios often require rapid assembly conclusions given "specified implant identifiers and specified abutment identifiers," with traceability of the basis for these conclusions to support processes such as selection, inventory preparation, warehousing, and surgical instrument preparation. In reality, data sources primarily consist of inventory parts master data and manufacturer connection information. These differ in specifications, coding, and update schedules, and the same part may have multiple version records. Missing or inconsistent compatibility information can easily lead to assembly uncertainty, further resulting in business consequences such as mismatch rework, delays, or difficulties in risk control.

[0003] In existing technologies, a common approach is to perform rule matching based on manufacturer manuals, lookup tables, or database fields: after standardizing attributes such as the connection type of parts, platform diameter, and screw specifications, connection is determined through equivalent matching or threshold matching; for generational differences, it usually relies on maintaining mapping tables or manually entering inheritance rules. In terms of system implementation, some solutions use multi-table join queries in relational databases, or abstract parts and specifications into entity relationship models for retrieval; other solutions, when data is relatively complete, verify each item in a preset judgment order and output matching results to guide selection and outbound approval.

[0004] The main shortcomings of the above solution in actual implementation are: when there are missing, conflicting, or uncovered new combinations of manufacturer connection data or inventory fields, the rule chain is difficult to continue, and the system can only give a static conclusion of "no match / not supported", which cannot describe the uncertainty brought about by the missing relationship; at the same time, the conclusion often lacks reproducible connection path basis, making it difficult to locate "where the break is and what relationship is missing", thus failing to support subsequent completion, review and closed-loop delivery, and the business side still needs to return to manual review and experience judgment.

[0005] Therefore, a method for determining implantation repair connectivity compatibility that can overcome the shortcomings of the existing technology is a problem that needs to be solved by those skilled in the art. Summary of the Invention

[0006] One objective of this invention is to propose a method for determining implant repair connection compatibility based on Node2Vec and compatibility knowledge graph retrieval. The core technical problem to be solved by this application is: under the constraints of only having inventory master data of parts and manufacturer connection information and the possibility of missing compatibility relationships, how to establish a traceable structured connection chain determination mechanism for a specified implant and abutment, and form verifiable candidate completion for missing relationships when a breakpoint occurs, so that the compatibility determination can be output in a closed loop and the basis can be recorded.

[0007] The implantation remediation connectivity compatibility determination method based on Node2Vec and compatibility knowledge graph retrieval according to embodiments of the present invention includes:

[0008] S1. Receive the specified implant identifier and the specified abutment identifier, and generate a compatibility knowledge graph based on the inventory parts master data and manufacturer connection information.

[0009] S2. Determine the connection chain template based on the compatibility knowledge graph. The connection chain template defines the connectability determination constraints in the order of connection form matching, platform diameter matching, screw specification matching, and series generational mapping. Search the connection chain from the specified implant node to the specified abutment node in the compatibility knowledge graph according to the connection chain template to obtain the search status. The search status is either a closed connection chain or a breakpoint description. The breakpoint description includes anchor nodes and missing relationship types.

[0010] S3. Based on the breakpoint description, perform Node2Vec representation learning on anchor nodes in the compatibility knowledge graph to obtain the node representation set. Determine the target node set based on the missing relation type. Calculate the representation similarity between anchor nodes and target nodes. If the representation similarity meets the similarity threshold, generate a candidate compatibility relation list. The candidate compatibility relation list includes candidate relations and candidate relation location identifiers.

[0011] S4. Based on the candidate compatibility relationship list, write the candidate relationships into the compatibility knowledge graph according to the candidate relationship position identifier to form an enhanced knowledge graph, and retrieve the verification connection chain in the enhanced knowledge graph according to the connection chain template;

[0012] S5. Determine the compatibility judgment result based on the retrieval status and the verification connection chain. When the retrieval status is a closed connection chain, determine the explicit compatibility judgment result and associate it with the closed connection chain. When the retrieval status is a breakpoint description and the verification connection chain is a closed connection chain, determine the inferred compatibility judgment result and associate it with the verification connection chain. When the verification connection chain is not a closed connection chain, determine the result that cannot be judged and associate it with the breakpoint description.

[0013] S6. Generate a connection compatibility determination record based on the compatibility determination result. The connection compatibility determination record contains the compatibility determination result and the closed connection chain, verification connection chain or breakpoint description associated with the compatibility determination result.

[0014] Optionally, S1 is as follows:

[0015] Receive a specified implant identifier and a specified abutment identifier, locate the implant part record and abutment part record in the inventory parts master data, and write the specified implant identifier and the specified abutment identifier into the corresponding node primary key respectively;

[0016] Extract the connection type, platform diameter, screw specification, and series generation fields from the inventory parts master data and manufacturer connection data, and map the field values ​​to preset enumeration values ​​to form a set of specification values ​​that can be used for relationship construction.

[0017] The implant component record and the abutment component record are used to generate implant node and abutment node, the specification value set is used to generate specification node, and the association relationship between component node and specification node is established to solidify the interface specification of the component.

[0018] Based on the manufacturer's connection data and the aforementioned relationships, connection form matching relationships, platform diameter matching relationships, screw specification matching relationships, and generational mapping relationships are generated. These four types of relationships are written into the relationship type set, and the starting node, ending node, and relationship type are recorded in the edge set. This results in a compatibility knowledge graph containing the node set, edge set, and relationship type set. Node encoding tables and relationship type encoding tables are also generated for one-hot encoding input in subsequent representation learning.

[0019] Optionally, S2 is as follows:

[0020] Based on the set of relationship types in the compatibility knowledge graph, four types of relationship types are identified: connection form matching relationship, platform diameter matching relationship, screw specification matching relationship, and series generation mapping relationship. The relationship type sequence is formed in the order of connection form matching relationship, platform diameter matching relationship, screw specification matching relationship, and series generation mapping relationship.

[0021] The relation type sequence is defined as the constraint segment sequence of the connection chain template, and each constraint segment is bound with the allowed retrieval relation direction and segment order position identifier to obtain the connection chain template with single path constraint.

[0022] Using a specified implant node as the starting retrieval node, the successor node that satisfies the corresponding relationship type is retrieved in the compatibility knowledge graph according to the first constraint segment of the connection chain template, and the successor node is determined as the first intermediate landing point.

[0023] Using the first intermediate landing point as the current node, continue to search and update the intermediate landing point according to the next constraint segment of the connection chain template, so that the search process forms an intermediate landing point corresponding to the segment order position identifier in each constraint segment.

[0024] When all constraint segments of the connection chain template are completed and the final intermediate landing point is the specified base node, the relation instances corresponding to each constraint segment are concatenated according to the constraint segment sequence to generate a closed connection chain, and the closed connection chain is determined as the retrieval state.

[0025] When a successor node cannot be found based on the next constraint segment during the retrieval process, the intermediate point of the last successful update is determined as the anchor node, the relation type corresponding to the next constraint segment is determined as the missing relation type, a breakpoint description containing the anchor node and the missing relation type is generated, and the breakpoint description is determined as the retrieval status.

[0026] Optionally, during the process of continuing to search and update intermediate landing points according to the next constraint segment of the connection chain template, with the first intermediate landing point as the current node, a template constraint reachability score is calculated for each candidate successor node obtained from the next constraint segment search, and an intermediate landing point for updating is selected and determined from the candidate successor nodes based on the template constraint reachability score. The template constraint reachability score is calculated by a constraint function, specifically:

[0027] ;

[0028] in, Indicates a candidate successor node Currently in the th When constraining segments, along the connection chain template The remaining constraint segment reaches the specified base node. Template constraint path count, This represents any node in the set of successor nodes obtained from the current constraint segment retrieval. This indicates the sequence number of the constraint segment that has been completed. Indicates a link template The sequence number variable of the middle constraint segment. This indicates the relationship between nodes generated from the node encoding table and nodes. The corresponding one-hot vector, This indicates the node generated by the node encoding table and the specified base station node. The corresponding one-hot vector, This represents the vector transpose operation. Indicates a link template The Middle Segment constraint segment relationship type, Indicates a link template The Middle The direction of the relationship between segment constraints and segment bindings. Indicates the relation type is And the direction is The relational adjacency matrix uses the node indices of the node encoding table as coordinates for its rows and columns. Matrix elements are set to one when an edge satisfying the relation type and direction constraints exists, pointing from a row node to a column node; otherwise, they are set to zero. Indicates by from arrive Perform matrix chain multiplication in the order described For connecting chain templates The total number of constraint segments included.

[0029] Optionally, S3 specifically refers to:

[0030] Read the breakpoint description to obtain the anchor node and the missing relationship type, and determine the segment position identifier corresponding to the missing relationship type according to the connection chain template;

[0031] Using the anchor node as the starting point of the walk, the first jump edge is selected in the compatibility knowledge graph according to the missing relation type constraint, and the subsequent jump edges are selected according to the subsequent relation type sequence constraint of the connection chain template, generating a walk sequence that is consistent with the order of the connection chain template;

[0032] Training samples are constructed based on the walking sequence. The central node and the context node are determined as node encoding features, and the relationship type from the central node to the context node is determined as the relationship type encoding feature.

[0033] The node encoding features are input into the node encoding neurons and represented using one-hot encoding. The relation type encoding features are input into the relation type encoding neurons and represented using one-hot encoding. The value of the relation type encoding features is determined by the missing relation type.

[0034] The node encoding features are mapped to node representations through the node embedding layer, the relation type encoding features are mapped to relation type representations through the relation type embedding layer, and the node representations and relation type representations are added together in the fusion layer to obtain the center node representation with missing relation type conditionalization.

[0035] The conditional center node representation is predicted by the context node prediction layer. Negative sampling is used to train and update the parameters of the node embedding layer, relation type embedding layer and context node prediction layer to obtain a set of node representations including anchor nodes.

[0036] The target node set is determined based on the relationship pointing range of the missing relationship type in the compatibility knowledge graph. The representation similarity between the anchor node and each target node in the target node set is calculated. A candidate compatible relationship list is generated when the representation similarity meets the similarity threshold. The candidate relationship in the candidate compatible relationship list consists of the anchor node, the missing relationship type, and the target node. The position identifier of the candidate relationship is determined by the segment position identifier.

[0037] Optionally, negative sampling is used to train and update the parameters of the node embedding layer, relation type embedding layer, and context node prediction layer. This includes: calculating the log loss value for each training sample based on the missing relation type conditionalized center node representation, the context node prediction layer output, and the negative sample nodes obtained from negative sampling; and performing parameter updates based on the log loss value. The log loss value is calculated by a loss function, which is specifically:

[0038] ;

[0039] in, This represents the logarithmic loss value for a single training sample. Represents the sigmoid activation function. This represents the central node output by the node embedding layer. The node represents a vector. This indicates the missing relation type output by the relation type embedding layer. Relational types represent vectors. This indicates the transpose operation. Represents the context node The output vector in the context node prediction layer. This indicates the number of nodes sampled according to their frequency distribution. One negative sample node Represents negative sample nodes The output vector in the context node prediction layer. This indicates the number of negative samples.

[0040] Optionally, S4 specifically refers to:

[0041] Read the candidate compatibility relationship list, group the candidate relationships according to their position identifiers, and obtain the candidate relationship subsets corresponding to each constraint segment of the connection chain template;

[0042] For each candidate relation, the anchor node, missing relation type, and target node are parsed, and the missing relation type is determined as the write relation type;

[0043] Based on the candidate relationship position identifier, the write relationship type is bound to the segment order position identifier of the corresponding constraint segment in the connection chain template, forming a one-to-one correspondence between the candidate relationship and the constraint segment;

[0044] Write candidate edges with anchor nodes as start nodes, target nodes as end nodes, and the type of relation to be written into the edge set of the compatibility knowledge graph, while keeping the original edge set in the compatibility knowledge graph from being replaced, to obtain the enhanced knowledge graph.

[0045] Based on the connection chain template, a structured retrieval is performed in the enhanced knowledge graph with a specified implant node as the starting retrieval node. During the retrieval process, only relation types and their corresponding candidate edges that are consistent with the segment order position identifier of each constraint segment are allowed to be used for each constraint segment.

[0046] During the retrieval process, intermediate landing points are updated segment by segment and relational instances are connected. When all constraint segments of the connection chain template are completed and the final intermediate landing point is the specified base node, a verification connection chain is generated.

[0047] The source of the relation instances used in each constraint segment in the verification connection chain is the original edge or candidate edge of the compatibility knowledge graph, and the output of the verification connection chain is used for compatibility determination.

[0048] Optional, S5 specifically includes:

[0049] Read the search status and determine whether the search status is a closed link chain. At the same time, read the verification link chain and determine whether the verification link chain is a closed link chain in the order of the link chain template.

[0050] When the retrieval status is a closed connection chain, the compatibility determination result is determined as an explicit compatibility determination result, and the closed connection chain is used as the associated connection chain of the compatibility determination result;

[0051] When the retrieval status is a breakpoint description and the verification connection chain is a closed connection chain, the compatibility determination result is determined as the inferred compatibility determination result, and the verification connection chain is used as the associated connection chain of the compatibility determination result;

[0052] When verifying that the connection chain is not a closed connection chain, the compatibility determination result is determined as an undeterminable result, and the breakpoint description is used as the associated breakpoint description of the compatibility determination result.

[0053] Optionally, reading the verification link and determining whether the verification link is a link closed in the order of the link template includes: calculating a closure determination value for the verification link and the link template, and determining whether the verification link is a link closed in the order of the link template based on the value of the closure determination value, wherein the closure determination value is calculated by a determination function, and the determination function is specifically:

[0054] ;

[0055] in, This represents the closure criterion value, which takes the value of... Indicates verification of the connection chain To connect the template A sequentially closed chain of connections takes the value of Indicates verification of the connection chain Not a closed link chain This indicates an indicator function that takes the value when the condition within the parentheses is true. Otherwise, the value is , Indicates verification of the connection chain The node sequence, Indicates the starting node. Indicates the final node, Indicates the specified implant node. Indicates the specified base station node. Indicates verification of the connection chain Number of relation instances Indicates a link template The number of constraint segments, Indicates verification of the connection chain The Middle The relation type of each relation instance, Indicates a link template The Middle Segment constraint segment binding relationship type, Indicates verification of the connection chain The Middle Directional encoding of relation instances, Indicates a link template The Middle Segment constraint segment binding relationship direction encoding, Indicates by from arrive Perform a series of multiplications.

[0056] Optionally, step S6 specifically includes:

[0057] Read the compatibility determination result, extract the compatibility determination result type and the closed connection chain, verification connection chain or breakpoint description associated with the compatibility determination result;

[0058] Write the specified implant node, specified abutment node, connection chain template, and compatibility determination result type into the main field of the connection compatibility determination record;

[0059] When the closed connection chain or verification connection chain associated with the compatibility determination result exists, the relation type, start node, end node, segment order position identifier, and relation instance source as the original edge or candidate edge of the compatibility knowledge graph corresponding to each constraint segment in the connection chain are written into the chain field of the connection compatibility determination record.

[0060] When the breakpoint description associated with the compatibility determination result exists, the anchor node, the missing relationship type, and the segment position identifier corresponding to the missing relationship type are written into the breakpoint field of the connection compatibility determination record to generate the connection compatibility determination record.

[0061] The beneficial effects of this invention are:

[0062] (1) This proposal proposes an improved method for determining the compatibility of implant-apron connections. By uniformly enumerating and writing information such as connection type, platform diameter, screw specifications, and generational sequence from the master data of inventory parts and the connection data of manufacturers into a compatibility knowledge graph, the determination order and directional constraints required for clinical assembly are solidified on the graph using a "connection chain template". This makes the compatibility retrieval no longer a loose attribute comparison, but a structured path determination that proceeds segment by segment along the template. In the retrieval process, the scheme introduces template constraint reachability scoring for candidate successor nodes, prioritizing intermediate landing points that can still reach the specified abutment in the remaining segment sequence, thereby reducing the accidental entry of uncertain paths when there are multiple branches or multiple matching edges in the data. When it is impossible to continue, the anchor node and the missing relationship type form a breakpoint description, directly locating "which segment of the template is missing and what relationship is missing", providing an interpretable mechanistic boundary for subsequent completion and review. Compared with the existing rule scheme that only returns "no match / no result", it is more conducive to business traceability and data governance closure.

[0063] (2) This proposal proposes a novel missing relation inference mechanism oriented towards breakpoints. Without introducing additional external data constraints, it transforms the uncertainty of incomplete connection information into a latent variable representation in the neighborhood of the graph structure: starting from the breakpoint anchor point, the missing relation type is used as the first-hop constraint, and a walking sequence is generated by strictly following the sequence of remaining relation types in the connection chain template. Random observation samples consistent with the assembly judgment order are constructed at the micro level. In representation learning, the missing relation type is used as an independent input and fused with relation type embedding and node embedding to obtain a node representation with "conditionalized missing relation type". Neighborhood prediction of the context is then achieved through negative sampling training. This design differs from embedding learning that is indiscriminate on the whole graph and simple similarity recall: the generation of its candidate relations is limited to the range of the missing relation type and has a segment position identifier. Therefore, the output is a list of candidate compatible relations that is writable, verifiable and consistent with the template segment position. This allows the key intermediate quantity - "whether the missing relation is possible" - to be repeatedly estimated under threshold control, rather than relying on manual experience to supplement.

[0064] (3) This proposal proposes a graph enhancement and verification method for a delivery-closed-loop approach. Candidate relations are appended to the knowledge graph according to their segment order and position, ensuring that existing edges are not replaced. The source of each relation instance is labeled as either an existing edge or a candidate edge. Subsequently, only relation types consistent with the template segment order and corresponding candidate edges are allowed for secondary structured retrieval. Verification connection chains are obtained, and the start and end nodes, segment count consistency, and segment-by-segment relation type and direction consistency are verified using a closure decision function. Based on the "original retrieval state + verification connection chain," three types of results are output: explicit compatibility, inferred compatibility, or indeterminate. The closed connection chain, verification connection chain, or breakpoint description is written into the connection compatibility determination record, which includes the start and end nodes, relation type, segment order position, and source of each relation instance. Thus, even under conditions of incomplete connection data, this method can provide a definite determination when a valid evidence chain exists, and a verifiable inference path or clear missing description when a breakpoint exists, enabling compatibility determination to form a deliverable closed-loop process from retrieval, inference, verification to record keeping. Attached Figure Description

[0065] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0066] Figure 1 The flowchart shows a planting remediation connectivity compatibility determination method based on Node2Vec and compatibility knowledge graph retrieval proposed in this invention.

[0067] Figure 2 This is a flowchart of the connection chain template retrieval method for a planting restoration connection compatibility determination method based on Node2Vec and compatibility knowledge graph retrieval proposed in this invention.

[0068] Figure 3 This is a flowchart of the breakpoint description and Node2Vec representation learning process for a planting repair connectivity compatibility determination method based on Node2Vec and compatibility knowledge graph retrieval proposed in this invention.

[0069] Figure 4 This is a flowchart of the enhanced knowledge graph and verification connection chain for a planting remediation connection compatibility determination method based on Node2Vec and compatibility knowledge graph retrieval proposed in this invention.

[0070] Figure 5 This is a flowchart illustrating the compatibility determination result of a planting repair connectivity compatibility determination method based on Node2Vec and compatibility knowledge graph retrieval proposed in this invention. Detailed Implementation

[0071] In Example 1, reference Figures 1 to 5A method for determining the connectivity compatibility of planting remediation based on Node2Vec and compatibility knowledge graph retrieval, comprising:

[0072] S1. Receive the specified implant identifier and the specified abutment identifier, and generate a compatibility knowledge graph based on the inventory parts master data and manufacturer connection information.

[0073] S2. Determine the connection chain template based on the compatibility knowledge graph. The connection chain template defines the connectability determination constraints in the order of connection form matching, platform diameter matching, screw specification matching, and series generational mapping. Search the connection chain from the specified implant node to the specified abutment node in the compatibility knowledge graph according to the connection chain template to obtain the search status. The search status is either a closed connection chain or a breakpoint description. The breakpoint description includes anchor nodes and missing relationship types.

[0074] S3. Based on the breakpoint description, perform Node2Vec representation learning on anchor nodes in the compatibility knowledge graph to obtain the node representation set. Determine the target node set based on the missing relation type. Calculate the representation similarity between anchor nodes and target nodes. If the representation similarity meets the similarity threshold, generate a candidate compatibility relation list. The candidate compatibility relation list includes candidate relations and candidate relation location identifiers.

[0075] S4. Based on the candidate compatibility relationship list, write the candidate relationships into the compatibility knowledge graph according to the candidate relationship position identifier to form an enhanced knowledge graph, and retrieve the verification connection chain in the enhanced knowledge graph according to the connection chain template;

[0076] S5. Determine the compatibility judgment result based on the retrieval status and the verification connection chain. When the retrieval status is a closed connection chain, determine the explicit compatibility judgment result and associate it with the closed connection chain. When the retrieval status is a breakpoint description and the verification connection chain is a closed connection chain, determine the inferred compatibility judgment result and associate it with the verification connection chain. When the verification connection chain is not a closed connection chain, determine the result that cannot be judged and associate it with the breakpoint description.

[0077] S6. Generate a connection compatibility determination record based on the compatibility determination result. The connection compatibility determination record contains the compatibility determination result and the closed connection chain, verification connection chain or breakpoint description associated with the compatibility determination result.

[0078] In this embodiment, step S1 specifically includes:

[0079] The designated implant identifier is recorded as The designated base station identifier is recorded as ,by and Accessing the inventory parts master data as the search key, the system locates implant part records and abutment part records respectively. When multiple part records with the same identifier exist in the inventory parts master data, a unique selection is performed based on the manufacturer identifier, series identifier, and generation identifier of the part record to obtain the implant part record and abutment part record used for mapping, and implant nodes are created for each. With base station nodes and will and Write the primary key field of the corresponding node so that subsequent link chain template retrieval can be performed using... and As a defined start and end point;

[0080] In the inventory parts master data and manufacturer connection information, extract the connection format field values ​​by field name and record them as follows: The platform diameter field value is denoted as The screw specification field value is recorded as follows: The generational field value is denoted as For the above four types of fields, an enumeration mapping table is established respectively, and the original field values ​​are normalized into preset enumeration values: for the platform diameter field value First, standardize the units of measurement, then round to the nearest decimal place to obtain discrete values, and map these discrete values ​​to platform diameter enumeration values. This process is then applied to the connection form field values. Screw specification field value , series generational field values The four types of enumeration values ​​are mapped to corresponding enumeration values ​​according to the encoding rules defined in the manufacturer's connection documentation, and then summarized into a specification value set. The set of specification values It includes enumerated values ​​for connection type, platform diameter, screw specification, and generation series, which are used for subsequent relationship construction and relationship type constraints.

[0081] A portion of the component node set is generated based on the implant component record and abutment component record, and based on the specification value set. Generate a portion of the specification node set, including implant nodes. With base station nodes Each part node is associated with a corresponding specification node, so that each part node has a traceable connection form, platform diameter, screw specification and series generation interface specification. The association relationship is used to solidify the part interface specification into the node neighborhood structure of the compatibility knowledge graph, so that the subsequent representation learning starting from the anchor node can cover the graph structure neighborhood related to the interface specification.

[0082] Based on the manufacturer's connection data and the aforementioned associations, four types of relationships are generated between part nodes: connection form matching relationship, platform diameter matching relationship, screw specification matching relationship, and series generation mapping relationship. For connection form matching relationship, platform diameter matching relationship, and screw specification matching relationship, the matching criterion is that the primary keys of the same specification nodes associated with the two part nodes are consistent. When a match is found, an edge with the first part node as the starting node and the second part node as the ending node is written, and an edge with the second part node as the starting node and the first part node as the ending node is also written. For series generation mapping relationship, directed edges are written according to the mapping direction between the source generation and the target generation recorded in the manufacturer's connection data, so that the connection chain template can advance in the direction of the series generation mapping constraint segment. The four types of relationships are written into the relationship type set and denoted as . And denote the edge set as The node set is denoted as follows: (The table lists the primary keys of the starting and ending nodes, the relationship type, and the node set as...) Thus, a compatibility knowledge graph is obtained. Depend on , and constitute;

[0083] To support the one-hot encoding input for subsequent Node2Vec representation learning, the generated node encoding table is denoted as follows: The relation type encoding table is denoted as Node encoding table Assign a unique consecutive integer index to each node's primary key; relation type encoding table. For a set of relation types Each relation type is assigned a unique consecutive integer index. When it is necessary to construct node encoding features, the total number of nodes and the node index are read, a one-hot vector with a length equal to the total number of nodes is generated, and the value of one is assigned to the corresponding position of the node index, while the value of zero is assigned to the other positions. When it is necessary to construct relation type encoding features, the total number of relation types and the relation type index are read, a one-hot vector with a length equal to the total number of relation types is generated, and the value of one is assigned to the corresponding position of the relation type index, while the value of zero is assigned to the other positions. This allows missing relation types to be directly input as relation type encoding features in the model computation chain during subsequent representation learning.

[0084] In this embodiment, step S2 specifically includes:

[0085] In obtaining a compatibility knowledge graph Then, from the compatibility knowledge graph Set of relation types The system reads the name field of each relation type and matches it with four preset relation type names to locate the four relation types, which are denoted as follows: Indicates a matching relationship in the form of a join. Indicates the matching relationship of platform diameters. This indicates the screw specification matching relationship. To represent a series of intergenerational mapping relationships, , , , Organized into relational type sequences according to the clinical assembly judgment order. The relation type sequence As the sole source of relational constraints for subsequent connection chain templates;

[0086] sequence of relation types Defined as a connection chain template The constraint segment sequence is defined, and each constraint segment is bound to the allowed retrieval relation direction and segment sequence position identifier, denoted as . Used to identify constraint segments in the connection chain template The sequential positions in the data are used to set the relationship direction for connection type matching, platform diameter matching, and screw specification matching to bidirectional searchable, while the relationship direction for generational mapping is set to unidirectional searchable. This unidirectional direction is consistent with the source-to-target generation direction recorded in the manufacturer's connection data. Each constraint segment is expressed as a triple. The triplet This includes relation type, relation direction, and segment order identifiers, thus obtaining a join chain template with a single path constraint. Composed of multiple Arranged in ascending order according to segment position identifiers;

[0087] The specified implant node is denoted as As the starting search node, initialize the current node as and read the connection chain template. Based on the first constraint segment and its relation type and direction, from the compatibility knowledge graph... edge set The system retrieves a set of successor nodes that satisfy the specified relationship type. When the relationship is bidirectionally searchable, the set of successor nodes includes nodes reachable from the current node along outgoing edges of the relationship type and nodes reachable along incoming edges of the relationship type. When the relationship is unidirectionally searchable, the set of successor nodes only includes nodes reachable from the current node along outgoing edges of the relationship type. This is to ensure the selection of intermediate nodes and the connection chain template. The subsequent constraint segments remain consistent, and the edge set is kept consistent. Construct a relation adjacency matrix based on relation type and relation direction, and then use a connection chain template. The remaining constraint segment is calculated for each candidate successor node to the specified base node, denoted as . The template constraint reachability score is calculated based on the constraint function.

[0088] ;

[0089] in, Indicates a candidate successor node Currently in the th When constraining segments, along the connection chain template The remaining constraint segment reaches the specified base node. Template constraint path count, This represents any node in the set of successor nodes obtained from the current constraint segment retrieval. This indicates the sequence number of the constraint segment that has been completed. Indicates a link template The sequence number variable of the middle constraint segment. This indicates the relationship between nodes generated from the node encoding table and nodes. The corresponding one-hot vector, This indicates the node generated by the node encoding table and the specified base station node. The corresponding one-hot vector, This represents the vector transpose operation. Indicates a link template The Middle Segment constraint segment relationship type, Indicates a link template The Middle The direction of the relationship between segment constraints and segment bindings. Indicates the relation type is And the direction is The relational adjacency matrix uses the node indices of the node encoding table as coordinates for its rows and columns. Matrix elements are set to one when an edge satisfying the relation type and direction constraints exists, pointing from a row node to a column node; otherwise, they are set to zero. Indicates by from arrive Perform matrix chain multiplication in the order described For connecting chain templates The total number of constraint segments included;

[0090] For each candidate successor node in the set of successor nodes of the first constraint segment calculate ,Will The candidate successor node with the largest value is determined as the first intermediate landing point. And will be used to reach The relation instance is bound to the segment order position identifier of the first constraint segment. When multiple candidate successor nodes have the same maximum... When retrieving values, the candidate successor node with the smallest value sorted by the node's primary key is determined as the first intermediate landing point. ;

[0091] With the first middle landing point As the current node, follow the connection chain template. The next constraint segment repeats the successor node search, and calculates the successor node for each candidate successor node set. , To determine the second intermediate landing point and the third intermediate landing point In each constraint segment, a relation instance is recorded to reach the corresponding intermediate landing point, so that the retrieval process forms an intermediate landing point that corresponds one-to-one with the segment sequence position identifier in each constraint segment, after completing the connection chain template. All constraint segments and the final intermediate point is the specified base node. At that time, the relation instances corresponding to each constraint segment are concatenated in order of segment position identifier to generate a closed connection chain. The closed connection chain Include along , , , Connect the node sequence and relation type sequence sequentially, and close the join chain. The search status has been confirmed.

[0092] When a subsequent node cannot be found based on the next constraint segment during the retrieval process, the connection chain template should be used. The propulsion position generates a breakpoint description. The intermediate point of the last successful update is designated as the anchor node and denoted as . The relation type corresponding to the next constraint segment is determined to be a missing relation type, denoted as . And the segment sequence position of the next constraint segment is marked as... Breakpoint description Includes anchor nodes Missing relation type Segment position identifier Describe the breakpoint Determine the search status and enable breakpoint description. The anchor nodes can serve as the input boundary for subsequent Node2Vec representation learning. Used to determine the starting point of the walk, and the type of missing relation. Used to limit the range of relationships pointed to by the target node set and the location where candidate relationships are written.

[0093] In this embodiment, step S3 specifically includes:

[0094] In one embodiment, reading the breakpoint description From the breakpoint description Extract anchor nodes Extract missing relation types and read the connection chain template. To determine the type of missing relation Corresponding segment position identifier Based on segment position identifier Positioning connection chain template Middle in missing relation type The subsequent constraint segments will be arranged in order of their relation types to form a sequence of subsequent relation types. This is used to impose sequential constraints on the jump edge relationship types of the walking sequence;

[0095] anchor node As the starting point for exploration, in the compatibility knowledge graph edge set Searching for missing relation types And satisfy the connection chain template The first jump edge set in the direction of the constraint segment relationship is selected using the Node2Vec second-order walk strategy. The Node2Vec second-order walk strategy is determined by the bounce parameter. With expansion parameters Control, bounce parameters With expansion parameters Given the preset configuration, after determining the first hop edge, the first hop node is obtained and used as the current node. The subsequent relation type sequence is then read sequentially. The relation type in the edge set The search query retrieves data that satisfies the current relation type and the join chain template. The set of subsequent jump edges corresponding to the constraint segment relationship direction is used to advance hop by hop until the sequence of subsequent relationship types is completed. If the subsequent jump edge set is empty, repeat the process a preset number of times. The walk generation process yields a set of walk sequences. The preset number of times Given a preset configuration, the set of walk sequences Each walk sequence in the algorithm contains a node sequence and a corresponding jump edge relationship type sequence, and the jump edge relationship type sequence satisfies the missing relationship type. The first hop appears and the subsequent relation type sequence is followed. Proceed in sequence;

[0096] Based on the set of walking sequences Construct a training sample set, using each jump edge in the walk sequence as a training sample, and determine the starting node of the jump edge as the center node and denote it as . The endpoint node of the jump edge is determined as the context node and denoted as... , will the central node With context node Through the node encoding table The data is converted into node-encoded features, which are represented using one-hot encoded vectors. The relation types of the training samples are limited to missing relation types. and missing relation types Through the relation type encoding table The features are converted to relational encoding features, which are represented using one-hot encoded vectors, thus ensuring that each training sample simultaneously carries the central node. Context nodes Types of missing relationships Three types of computable inputs;

[0097] A single hidden layer neighborhood prediction structure is constructed and trained. The node embedding layer is set as a trainable embedding matrix lookup table structure, mapping the node encoded features to a dimension of 1. The nodes represent vectors, with dimensions... Given the preset configuration, the relation type embedding layer is set as a trainable relation type embedding matrix lookup table structure, mapping the relation type encoded features to a dimension of [missing information]. The relation type represents a vector, and the central node is used in the fusion layer. Node representation vectors and missing relation types The relationship type representation vectors are summed to obtain the center node representation vector for conditionalizing missing relationship types. The context node prediction layer is set as an output vector lookup table structure, maintaining a dimension for each candidate context node. The output vector is then used, and negative sampling is employed to train and update the parameters of the node embedding layer, relation type embedding layer, and context node prediction layer. The number of negative samples is denoted as . negative sampling quantity Given by a preset configuration, based on a set of walk sequences The frequency of each node as a context node is counted and normalized to obtain the node occurrence frequency distribution. This node occurrence frequency distribution is then used to analyze the compatibility knowledge graph. Node set Negative sample nodes are extracted from the training data. For each training sample, the inner product value is mapped to a probability value using a sigmoid activation function, and log loss is used as the optimization objective. The log loss is calculated according to the loss function:

[0098] ;

[0099] in, This represents the logarithmic loss value for a single training sample. Represents the sigmoid activation function. This represents the central node output by the node embedding layer. The node represents a vector. This indicates the missing relation type output by the relation type embedding layer. Relational types represent vectors. This indicates the transpose operation. Represents the context node The output vector in the context node prediction layer. This indicates the number of nodes sampled according to their frequency distribution. One negative sample node Represents negative sample nodes The output vector in the context node prediction layer. Indicates the number of negative samples;

[0100] The parameters of the node embedding layer, relation type embedding layer, and context node prediction layer are updated using backpropagation based on the logarithmic loss to obtain the anchor node. The nodes within represent sets. , where nodes represent sets any node The node representation vector is denoted as ;

[0101] Based on missing relation type In compatibility knowledge graph The relationship in the middle points to the range that determines the target node set. In the edge set The search relation type is a missing relation type. And satisfy the connection chain template For all edges corresponding to the constraint segment relationship direction, the primary keys of the endpoint nodes of these edges are summarized and deduplicated to obtain the target node set. For anchor nodes With the target node set The similarity of each target node is calculated. The similarity is obtained by combining the cosine similarity with the inner product of the node representation vectors and the L2 norm. The similarity threshold is then read. Similarity threshold Given by a preset configuration, when the similarity meets a similarity threshold Under the condition, anchor node Missing relation type The target nodes are combined to generate candidate relationships, and the segment order positions are marked. Write the candidate relation position identifier, sort the candidate relations that meet the threshold in descending order of representation similarity, and truncate the first few relations. A list of candidate compatibility relationships is formed, among which This is the upper limit of the number of candidate relationships in the candidate compatibility relationship list, and is given by the preset configuration.

[0102] In this embodiment, step S4 specifically includes:

[0103] Read the candidate compatibility relationship list, and denote the candidate compatibility relationship list as... Each candidate relationship includes an anchor node, a missing relationship type, and the location identifiers of the target node and the candidate relationship. This forms the basis of the candidate compatible relationship list. Parse the position identifiers of each candidate relation, group them according to the position identifiers, and denoted as the set of candidate relation subsets. ,in The values ​​representing the segment position identifiers, and the candidate relation subsets. For corresponding connection chain templates The middle section sequence position is marked as Constraint segment;

[0104] For the subset of candidate relations Each candidate relation in the table is subjected to field decomposition to obtain the anchor node, denoted as . The missing relation type is denoted as The target node is denoted as and missing relation types Determined to be a write relation type, read the join chain template. The constraint segment sequence is based on the segment sequence position identifier corresponding to the candidate relation position identifier. Locate the target constraint segment and verify the relationship type bound to the target constraint segment and the write relationship type. Consistency is established by creating a one-to-one correspondence between candidate relations and target constraint segments, ensuring that write operations occur within the connection chain template. The segment position identifier dimension remains traceable;

[0105] In the compatibility knowledge graph, it is denoted as The set of edges is denoted as The above performs an append write, using the anchor node of each candidate relationship. As the starting node and the target node As the endpoint node, written to the relation type As a relation type, the generated candidate edge set is denoted as... , is the set of candidate edges Each candidate edge in the algorithm is written with a segment order position identifier attribute, the value of which is the segment order position identifier in the candidate relation position identifier corresponding to that candidate edge. This enables candidate edges to possess segment order semantics that can be filtered by constrained segments, maintaining compatibility with knowledge graphs. Central Plains has a set of borders If it is not replaced, the enhanced knowledge graph will be denoted as And will enhance the knowledge graph The set of edges is denoted as The set of edges From the original set of edges With candidate edge set Composed of, and also a set of edges Each edge in the graph is written with a source label. The source label value is either an existing edge or a candidate edge in the compatibility knowledge graph, which is used to verify the source label of the connection chain in the future.

[0106] Based on connection chain template In enhancing knowledge graphs Perform a structured search and read the specified implant node, denoted as... The specified base station node is denoted as Initialize the current node to and according to the connection chain template The segment position identifiers are incremented and progressively advance to the next segment. A segment constraint segment allows the retrieval of successor nodes only using the relation type and relation direction bound to it, and applies a segment order position identifier filtering condition to candidate edges, ensuring that the used candidate edges satisfy the condition that the segment order position identifier attribute of the candidate edge is equal to... Before each stage of advancement, a set of backward reachable nodes is constructed to constrain the selection of intermediate landing points: the first... The last constraint segment is considered as the remaining constraint segment. Based on the relation type and relation direction of the remaining constraint segment, starting from the specified base node... Starting from the beginning, expand backward segment by segment to obtain the set of nodes reachable in the backward direction, denoted as . During the reverse expansion process, the segment order position identifier filtering condition is also applied to the candidate edges, so that the candidate edges used satisfy the condition that the segment order position identifier attribute of the candidate edge is consistent with the segment order position identifier corresponding to the reverse expansion. After obtaining the set of successor nodes through segment constraint retrieval, only the set of nodes that are reachable from the back is retained. The successor node is taken as the set of available successor nodes, and the next intermediate landing point is selected from the set of available successor nodes by sorting the nodes by their primary keys. The relation instance used to reach the next intermediate landing point is added to the relation instance sequence.

[0107] During the retrieval process, intermediate landing points are updated segment by segment, and relationship instances are linked together. Once the connection chain template is completed... All constraint segments and the final intermediate point is the specified base node. At that time, the sequence of relation instances and the sequence of nodes are combined to generate a verification connection chain, denoted as . In verifying the connection chain The source annotation of the relation instance used for each constraint segment is provided by the augmented knowledge graph. edge set The source annotation field of the corresponding edge is determined. The source annotation value is either an existing edge or a candidate edge in the compatibility knowledge graph. The verification connection chain containing the source annotation is then established. The output is used for subsequent compatibility determination.

[0108] In this embodiment, step S5 specifically includes:

[0109] The read search status is denoted as The search status The results are output from the structured retrieval stage and saved in record form. The retrieval status type is either a closed join chain or a breakpoint description. Simultaneously, the verified join chain is read and recorded as... The verification connection chain The structured retrieval phase of the enhanced knowledge graph outputs and saves it in record form. The read link template is denoted as... Connecting chain template Contains a sequence of constraint segments arranged in ascending order of segment position identifiers; the specified implant node is read and denoted as... The specified base station node is denoted as , used for checking the start and end nodes during closure determination;

[0110] Verify the connection chain Parsing into a sequence of nodes and a sequence of relation instances, to verify the connection chain. Parse to obtain the node sequence sequence of relation instances , of which Instance of Relationships Connecting nodes With nodes and from relation instances Read relation type to get From the connection chain template The number of constraint segments read is denoted as and read the first The relationship type of segment constraint segment binding is obtained Each relation instance The starting node in the graph edge record is denoted as... The endpoint node is denoted as and based on and Consistency calculation relation instance direction encoding :when and season ,when and season Connect the chain template The Middle The direction encoding of segment constraint segment binding relationship is denoted as When the constraint segment allows bidirectional retrieval, it means... When the constraint segment only allows retrieval along the edge record direction, let ;

[0111] Verify the connection chain Perform connection chain determination according to the connection chain template sequence, and use the closure determination value. Output the judgment result, first perform the prerequisite check: when or or At that time, directly Set as Then, the closure determination is completed. If the prerequisite checks pass, the consistency of the relation type and relation direction of each segment's position identifier is checked segment by segment, and the closure determination value is calculated according to the determination function. :

[0112] ;

[0113] in, This represents the closure criterion value, which takes the value of Indicates verification of the connection chain To connect the template A sequentially closed chain of connections takes the value of Indicates verification of the connection chain Not a closed link chain This indicates an indicator function that takes the value when the condition within the parentheses is true. Otherwise, the value is , Indicates verification of the connection chain The node sequence, Indicates the starting node. Indicates the final node, Indicates the specified implant node. Indicates the specified base station node. Indicates verification of the connection chain Number of relation instances Indicates a link template The number of constraint segments, Indicates verification of the connection chain The Middle The relation type of each relation instance, Indicates a link template The Middle Segment constraint segment binding relationship type, Indicates verification of the connection chain The Middle Directional encoding of relation instances, Indicates a link template The Middle Segment constraint segment binding relationship direction encoding, Indicates by from arrive Perform a series of multiplications;

[0114] Search status Perform type determination: When the search status type is a closed join chain, start from the search status The closed link obtained from the parsing is denoted as The closed connection chain Includes node sequence and connection chain template The sequence of relational instances arranged in segment order, with the compatibility determination result denoted as... and the compatibility determination result The result type field is written to the explicit compatibility judgment result, and the closed join chain is also included. Write compatibility determination results Set the associated connection chain field and empty the associated breakpoint description field;

[0115] When the retrieval status type is breakpoint description, from the retrieval status The breakpoint description obtained from the analysis is denoted as The breakpoint description Includes anchor nodes, missing relation types, and segment position identifiers, based on closure judgment values. Verify the connection chain Classify and process: When At that time, the compatibility determination result will be... The result type field is written into the inference compatibility decision result, and the verification connection chain is also included. Write the associated connection chain field, and also include the breakpoint description. Write the associated breakpoint description field when At that time, the compatibility determination result will be... Write "Unable to determine result" in the result type field, set the associated link field to null, and specify the breakpoint description. Write the associated breakpoint description field to include the compatibility judgment result. Write to the result storage table, which must contain at least a result type field, an association link chain field, and an association breakpoint description field.

[0116] In this embodiment, step S6 specifically includes:

[0117] The compatibility determination result is recorded as Compatibility determination result It includes the compatibility determination result type and the associated object. The associated object takes the value of a closed connection chain under different compatibility determination result types. Verify the connection chain Or breakpoint description Based on the compatibility determination results The type of compatibility determination result extracted is denoted as Compatible judgment result types The values ​​are: explicit compatibility determination result, inferred compatibility determination result, and indeterminate result. Simultaneously, the specified implant node is read and recorded as... The designated base station node is denoted as The connection chain template is denoted as Connecting chain template Contains a sequence of constraint segments arranged in ascending order of segment position identifier;

[0118] The generated connection compatibility determination record is denoted as And record the connection compatibility determination. The primary field is denoted as The chain field is denoted as The breakpoint field is denoted as , will specify implant node Specify base station node , Connecting chain template Compatibility determination result type Write to main field Connecting chain template During writing, the connection chain template will be used. Each constraint segment is extracted as a record item. Each record item contains a segment sequence position identifier and the relationship type bound to that constraint segment, making the main field... Able to reproduce the connection chain template The semantics of segmentation;

[0119] When the compatibility determination result Related closed chain Existence or verification of the connection chain If the connection chain exists, it is parsed into a sequence of nodes and a sequence of relation instances, and written into the chain field. The node sequence of the connected chain is denoted as ,in Indicates the first 1 node The segment sequence number is used to denote the sequence of relational instances of the connection chain. ,in Indicates the first The segment constrains the relation instance corresponding to the segment, and the relation instance... Connecting nodes With nodes The number of constraint segments in the connecting chain is denoted as Number of constraint segments From the sequence of relation instances The number of elements is obtained directly from each relation instance. Read the relation type and record it as And record the segment position identifier as ,in Values The relation type corresponding to each constraint segment Starting point node End point Segment position identifier Write chain field The The chain record, the first Chained records are stored sequentially using fixed fields, making the chain fields... It can replay the link according to the segment position;

[0120] Write the relation instance source to the chain field. When the associated objects are a closed join chain At that time, the source of the relation instance recorded in each chain is written into the original edge of the compatibility knowledge graph. When the associated object is a verification connection chain... At that time, from the verification connection chain Each relation instance Read the source annotation and record it as Source labeling The value is either an existing edge or a candidate edge in the compatibility knowledge graph, and the source is labeled. Write to chain field The Chained records, making the chain field Capable of identifying and verifying connection chains segment by segment. Sources of examples of relationships between China and other countries;

[0121] When the compatibility determination result Related breakpoint descriptions When it exists, parse the breakpoint description. The anchor node is denoted as The missing relation type is denoted as The segment position identifier corresponding to the missing relation type is denoted as and anchor node Missing relation type Segment position identifier Write breakpoint field When the breakpoint description If it does not exist, set the breakpoint field. Empty, when the connection chain is closed With verification connection chain If neither exists, then the chain field will be used. Set to null, complete the main field. Chain field With breakpoint field After writing, a connection compatibility determination record is generated. And record the connection compatibility determination. Write to the connection compatibility determination record table.

[0122] Example 2:

[0123] This embodiment runs on a Linux server environment, with the operating system Ubuntu 22.04 LTS and Python version 3.10. NetworkX is used for graph structure construction and structured retrieval, NumPy and SciPy are used for numerical computation, and PyTorch is used for representation learning and training. When the target set is large and similarity retrieval needs to be accelerated, FAISS is used. To ensure the reproducibility of the results, the random seed seed=42 is fixed, and all key hyperparameters are written into the same configuration file, including the embedding dimension d, negative sampling number K, number of walks N_walk, walk length L, similarity threshold τ, number of candidate truncations TopK, number of training epochs, and learning rate lr. The same input yields deterministic output under the same configuration.

[0124] The data source consists of two parts: inventory parts master data D_inv and manufacturer connection data D_mfg. The record granularity of D_inv is part record, and the fields include part identifier, manufacturer, series, generation, connection type, platform diameter, screw specifications, etc. D_mfg provides interface specification, series and generation mapping direction and known connection relationship. The input of the judgment task is a specified implant identifier ID_I and a specified abutment identifier ID_A. The system first locates the corresponding part record in D_inv. When there are multiple records for the same identifier, a master record is selected according to the uniqueness strategy. The uniqueness sorting key consists of manufacturer consistency, series consistency, generation consistency and record update time (or version number) to ensure the stability of subsequent mapping and retrieval results.

[0125] To unify the differences between D_inv and D_mfg in terms of caliber, encoding, and units, this embodiment standardizes and enumerates four types of fields: connection type, platform diameter, screw specification, and series generation. The connection type is mapped to the enumeration code ConnType=CT_01, the screw specification is mapped to the enumeration code Screw=SC_M18, the series generation is mapped to the enumeration code Gen=G3, and the platform diameter is uniformly unitized in mm and discretized according to the configured decimal place rules. The discretized diameter is denoted as d_plat and then mapped to the enumeration code Plat=PD_45. After standardization, each part record is fixed as an interface quadruple (ConnType, Plat, Screw, Gen). This quadruple serves as the unified input caliber for subsequent relation construction, template retrieval, and breakpoint inference.

[0126] The experimental evaluation dataset is constructed with a set of sample pairs S={(ID_I, ID_A)}. The compatibility label y comes from traceable sources: manufacturer comparison table, structured entries in manufacturer manuals, and historical assembly success records. To evaluate the breakpoint completion capability, a missing scenario data version D_mfg_mask is constructed: without changing the true compatibility of the samples, a certain type of relational information is masked according to the configuration rules. The masking process is controlled by seed and a masking list is recorded for subsequent statistical analysis of breakpoint rate, candidate hit rate (Hit@K), and re-examination closure rate, ensuring the reproducibility of comparative experiments and ablation experiments.

[0127] In this embodiment, the inventory parts master data D_inv and the manufacturer connection data D_mfg are uniformly abstracted into a compatibility knowledge graph G=(V,E,T), where V is the set of nodes, E is the set of edges, and T is the set of relation types. The schema of the graph consists of two types of nodes: parts nodes and specification nodes. Part nodes include two categories: implant parts and abutment parts. The primary key of the node uses the part identifier PartID (which may be combined with manufacturer / series / generation to form a unique key if necessary). Specification nodes are used to carry the part interface specifications. The values ​​of specification nodes come from the completed enumeration results, including connection form ConnType=CT_**, platform diameter Plat=PD_**, screw specification Screw=SC_**, and series generation Gen=G*. In implementation, specification nodes can be used as a unified Spec node type and distinguished by the field SpecType∈{ConnType,Plat,Screw,Gen}, or they can be split into four types of nodes. However, this embodiment uses a unified Spec node to simplify relation management and retrieval.

[0128] The graph's set of relationship types T contains at least four core component relationships: connection form matching relationship R_conn, platform diameter matching relationship R_plat, screw specification matching relationship R_screw, and series generation mapping relationship R_genmap. Among them, R_conn, R_plat, and R_screw are used to express the consistency of two components in the corresponding interface specifications, which can be regarded as bidirectional availability according to business semantics. R_genmap is used to express the connection inheritance / conversion relationship between different generations of the series, and the direction is defined as a directed edge according to the manufacturer's data. In addition to the above four types of component relationships, the graph also contains the fixed relationship from component to specification, which is used to fix the interface specification to the component's neighborhood. The relationship type can be uniformly recorded as R_hasSpec, and its corresponding specification dimension is identified by the edge attribute SpecType.

[0129] The graph construction process is executed in three steps. First, node generation: Traverse the part records in D_inv that are within the decision range, creating a part node for each record. Simultaneously, summarize all enumerated specification values ​​and create a specification node for each different specification value, ensuring that the same enumeration value appears only once in the graph. Second, establishing part-specification relationships: For each part node, read its standardized interface quadruple (ConnType, Plat, Screw, Gen), connect it to the corresponding four specification nodes, and write edges (Part -> Spec, type=R_hasSpec, SpecType=...), making the part interface specifications traceable and reusable in the graph structure. Third, deriving four types of core relationship edges between parts: For any two part nodes u and v, if their connection form enumeration values ​​are the same, write an R_conn edge; if their platform diameter enumeration values ​​are the same, write an R_plat edge; if their screw specification enumeration values ​​are the same, write an R_screw edge. For the three types of matching relationships mentioned above, bidirectional edges are written to support the bidirectional constraints of template retrieval. For the series of generational mapping relationships R_genmap, directed edges (u -> v, type=R_genmap) are written according to the source generation to target generation mapping direction explicitly given in D_mfg, and attribute fields such as manufacturer source and mapping rule number are retained to facilitate subsequent traceability and auditing. After completion, a compatibility knowledge graph G is obtained, and R_conn, R_plat, R_screw, and R_genmap are used as the set of available relation types for subsequent connection chain templates.

[0130] To support subsequent representation learning and matrix computation, this embodiment generates two encoding tables after graph construction. The node encoding table M_V assigns a continuous integer index idx_v to each node's primary key, which is used to map the node to the vector space or the row and column coordinates of the adjacency matrix. The relation type encoding table M_T assigns a continuous integer index idx_t to each relation type, which is used to map the relation type to a trainable relation type embedding or to determine relation type constraints. The encoding tables maintain an append-only expansion rule when the graph is updated (adding candidate edges), that is, the existing indexes remain unchanged, and only new indexes are added for newly appearing nodes or relation types, thereby ensuring the consistency of indexes and the reproducibility of results during the training and review stages.

[0131] In this embodiment, a connection chain template for assembly determination is defined on the compatibility knowledge graph G. The template consists of relation constraints with a fixed segment order, from pos=1 to pos=4: connection form matching R_conn, platform diameter matching R_plat, screw specification matching R_screw, and generational mapping R_genmap. Each segment constraint consists of three items: relation type rel, allowed direction dir, and segment order identifier pos. Among them, R_conn, R_plat, and R_screw allow bidirectional retrieval, and the direction identifier is denoted as dir=bi. R_genmap only allows retrieval according to the manufacturer-defined direction, and the direction identifier is denoted as dir=uni. Through this template, the compatibility determination is constrained to be a structured path retrieval that must be satisfied sequentially according to the segment order, rather than an unordered attribute comparison.

[0132] The structured search starts with the specified implant node I and ends with the specified abutment node A, proceeding segment by segment according to the template sequence. The search process maintains the current node cur and the sequence of matched relation instances chain. In the pos segment, the system only allows the use of the relation type rel specified in the segment, and obtains the successor node set Next(cur, rel, dir) from cur according to the dir constraint: when dir=bi, the successor set contains the out-neighbors and in-neighbors of cur for this relation; when dir=uni, it only contains the out-neighbors of this relation. If Next is empty, it means that the search cannot continue in this segment, the search stops immediately and a breakpoint description is generated.

[0133] When there are multiple candidate successor nodes for a certain segment, this embodiment can enable template reachability scoring to select intermediate landing points, thereby reducing the probability of entering invalid paths in a multi-branch graph structure. The scoring idea is to prioritize candidate nodes that are still more likely to reach the endpoint A under the constraints of the remaining segment order of the template. In implementation, the score can be used as one of the intermediate landing point selection rules: if the score is enabled, the node with the largest score among the candidate successor nodes is selected as the next intermediate landing point; if the score is not enabled, a deterministic rule (such as sorting by the node primary key and taking the minimum) can be used as the landing point selection method to ensure that the results are reproducible. Regardless of which selection strategy is adopted, the retrieval must strictly follow the segment order of the Template and record the specific edges used in each segment (start point, end point, relation type, direction, segment order) so that an interpretable evidence chain can be output later.

[0134] The retrieval output is divided into two categories. The first category is a closed connection chain: when all four segments pos=1..4 are completed and finally reach node A, a sequence of relation instances chain arranged in segment order is obtained, which together with the corresponding node sequence forms a closed connection chain ClosedChain. This closed connection chain is the evidence chain for explicit compatibility judgment, which can explain segment by segment why I and A can be connected. The second category is a breakpoint description: when the retrieval cannot find any successor node in a certain segment, the node successfully reached in the previous segment is recorded as the anchor node, the relation type required in the current segment is recorded as the missing relation type missing_rel, and written into the current segment order pos to form a breakpoint description Breakpoint=(anchor, missing_rel,pos). The breakpoint description is used to clearly indicate which segment of the template is broken in and what relation is missing, and serves as the input boundary for subsequent breakpoint completion and verification review.

[0135] The following pseudocode implements the connection chain template retrieval function `template_search(G, Template, I, A, useScore)`, used to determine whether a closed evidence chain can be formed between the implant node `I` and the abutment node `A` in a fixed segment order within the knowledge graph `G`. The algorithm starts from `I` and sequentially executes each relation constraint `(rel_pos, dir_pos)` in the template. In the neighborhood of the current node `cur`, it only searches for the next-hop candidate set `Cand` that satisfies the relation type and direction constraints. If a segment `Cand` is empty, it immediately returns the breakpoint `bp`, which includes the anchor node `anchor=cur`, the missing relation type `missingRel=rel_pos`, and the breakpoint segment order `pos`, used for subsequent breakpoint completion. If there is more than one candidate and `useScore` is enabled, `TemplateReachScore` prioritizes the intermediate landing point that is more likely to reach `A` under the remaining segment order of the template, and uses the deterministic rule `TieBreakByMinID`. Eliminate the uncertainty of parallel nodes; otherwise, select the next node with the smallest ID. Record a segmented edge instance to `chain` for each segmented step. When all segments are completed, if the final node is equal to `A`, return the closed chain as explicit evidence; otherwise, return the end breakpoint, indicating that although the segmented steps have been completed, the target node has not been reached.

[0136]

[0137] When the output is a breakpoint description Breakpoint=(anchor, missing_rel, pos), this embodiment starts the breakpoint completion process. First, the breakpoint segment sequence pos is located according to the connection chain template Template. The remaining constraint sequence Remain=[(rel_pos,dir_pos), (rel_{pos+1},dir_{pos+1}), ...] from the beginning of the template to the end of the last segment is extracted, where rel_pos is missing_rel. Then, a constrained random walk is performed with the anchor node as the starting point of the walk: the first hop only allows the selection of edges with relation type missing_rel and direction satisfying dir_pos. From the second hop onwards, the edges corresponding to relation type and direction are selected in strict accordance with the segment sequence of Remain until the remaining segment sequence is completed or there are no available edges in a certain hop. N_walk walk sequences are generated repeatedly. Each sequence maintains the consistency of the first hop = missing relation and subsequent hops = the remaining segment sequence of the template, so that the sampled neighborhood context is aligned with the structural order of the assembly decision, instead of walking in the whole graph without constraints.

[0138] In the representation learning phase, this embodiment trains conditional node representation vectors. For each node v, a node embedding z_v is maintained, and for the missing relation type missing_rel, a relation type embedding r_missing is maintained. The embedding dimension is d. The training samples are generated from the above-mentioned constraint walk sequence, using the Skip-gram approach: the starting point of a hop is taken as the center node center, and the ending point is taken as the context node context. The conditional representation of the center node is h = z_center + r_missing, that is, the missing relation type condition is explicitly injected into the center node representation. Context prediction is trained using negative sampling. Negative samples are extracted from the node frequency distribution. The number of training epochs is 1, the learning rate is lr, and the number of negative samples is K. After training, a representation set {z_v} containing the anchor point and its neighboring nodes is obtained. This representation reflects the structural similarity under the missing_rel condition and template order constraint.

[0139] In the candidate relation generation stage, this embodiment does not perform indiscriminate recall across the entire graph. Instead, it first constructs a target node set TargetSet, which is obtained by summing all edge endpoint nodes in the graph that satisfy the relation type missing_rel and the direction dir_pos. This ensures that candidates only fall within the allowed pointing range of the missing relation. Then, it calculates the similarity between the anchor point and the target node sim(anchor, t) = cosine(z_anchor, z_t), filters target nodes with sim >= τ, and selects the top K nodes in order of similarity from high to low to generate a candidate relation list CandList. Each candidate in the list contains four fields: (anchor, missing_rel, target, pos), where pos is used to identify which part of the template the candidate relation should be written back to, ensuring that it is not misused across segments during subsequent review.

[0140] In the graph enhancement and verification stage, this embodiment appends candidate edges from CandList to the knowledge graph to form an enhanced graph G+. During writing, existing edges are not overwritten. All newly added edges are marked with the attributes source="candidate" and pos=pos, while existing edges default to source="original". Subsequently, the same template retrieval is performed on G+ to obtain the VerifyChain, but with an added constraint: in the pos segment, only candidate edges with the same pos attribute and segment order are allowed to be used, preventing candidate edges from being incorrectly used in other segments. Other segments still use the original edges or candidate edges from the same segment according to the template constraint. The retrieved VerifyChain will carry edge source annotations (original / candidate) segment by segment to explain the inference basis.

[0141] Finally, this embodiment performs a closure check on VerifyChain to form a verification loop: it checks whether the starting point is the specified implant node I, whether the ending point is the specified abutment node A, and whether the number of chain segments is equal to the number of template segments m. It also checks whether the relationship type and direction are consistent with the corresponding segment of the Template. If the closure check passes, it is considered that the verification connection chain is closed according to the template, and the compatibility conclusion can be inferred. If the closure check fails, the breakpoint description is retained and the conclusion that it cannot be determined is entered.

[0142] The first pseudocode snippet below, `infer_candidates(G, Template, bp, Params)`, is used to infer "missing relation completion candidates" for the generated breakpoints. After inputting the breakpoint `bp=(anchor, missingRel, pos)`, the algorithm first extracts the remaining segment sequence `Remain` from the template starting at `pos`, and performs a constrained random walk starting from `anchor`: the first hop must be on `missingRel`, and subsequent hops strictly follow the segment sequence and direction of `Remain`, thus aligning the sampling context with the template structure. Based on these walk sequences, a conditional representation `Z` (node ​​embedding, introducing relation conditions for missing relations) is trained. The candidate space is then restricted to the target node set `TargetSet` that satisfies the direction constraint of `missingRel`. The cosine similarity between the anchor and the target is calculated, and after filtering by a threshold `τ`, the `TopK` nodes with the highest similarity are selected. The candidate edge list `CandList=(anchor, missingRel, target, pos, sim)` is output, where `pos`... Clearly identify which section of the template the candidate belongs to, so that it can be used by section during subsequent verification;

[0143] The second pseudocode segment `recheck_on_augmented_graph(G, Template, I, A, CandList)` is used to transform the "inference" into "verifiable evidence". The algorithm appends `CandList` to the augmented graph `G+`, adding edge labels `src=cand`, `pos`, and `sim` without overwriting the original edges. Then, it searches segment by segment again from `I` using the same template, but only two types of edges are allowed at each segment order `pos`: the original edge `src=orig` and the candidate edge `src=cand` that matches the segment order and `e.pos=pos`, to avoid misuse of candidate edges across segments. If no available edge is available in any segment, the recheck fails and returns an unclosed result. If each segment is successfully traversed, a verification chain `VerifyChain` is obtained. Finally, `IsClosedByTemplate` is used to verify the consistency between the start and end points and the segment order. Only when the verification is true is the closed result output, thus ensuring that the inference compatibility must meet the verification closed-loop requirement of "closing the loop according to the template".

[0144]

[0145]

[0146] The output of this method is not a simple compatibility / incompatibility statement, but rather three categories of conclusions based on the completeness of the evidence. This makes it easier for the business side to understand and verify the conclusions, and also facilitates subsequent closed-loop iterations (supplementing dictionaries, relationships, and rules).

[0147] 1. Explicit compatibility

[0148] Executing template_search(G, template, implant, abutment) on the original knowledge graph G, if the steps are followed in a fixed order of "connection form → platform diameter → screw specification → generational mapping" and finally reach the base node, it is directly determined to be explicit compatibility. At this time, the output evidence is a closed connection chain, and each segment of the chain can correspond to the real relation edge in the graph (source is original), so it has the strongest interpretability.

[0149] 2. Inference Compatibility

[0150] If the original search fails at a certain point, it will output a breakpoint bp=(anchor, missingRel, pos), indicating that a relation like missingRel is missing at segment pos. In this case, proceed to:

[0151] The `infer_candidates()` function performs constrained walks starting from the anchor point and trains conditional embeddings to generate a list of candidate edges for missing relationships (with segment order `pos` and similarity).

[0152] Write the candidate edges into the enhanced graph G+ in an append manner (without overwriting the original edges, and label the candidate edges with source=candidate).

[0153] If recheck_on_augmented_graph() is executed on G+ and a closed chain can be retrieved according to the same template segment order, it is determined to be inference compatible.

[0154] The key point of the conclusion is that closure must be checked again; that is, the inferred candidate edges must be able to close the entire template chain. Otherwise, the inference compatibility is not output to avoid misjudgment based solely on similarity.

[0155] 3. Undetermined

[0156] If the original search breakpoint plus enhanced re-examination still cannot close the gap according to the template, the output will be "Undetermined". This type is not equivalent to incompatibility, but rather indicates that the current map evidence is insufficient or the breakpoint completion fails to provide a verifiable closure chain. In business terms, it can proceed to manual review or data completion process.

[0157] To ensure that each decision is traceable and reviewable, this method generates a decision record. The record must cover not only the closed chain but also the breakpoints and candidate completion information, which facilitates error analysis (which segment is broken, what type of relationship is missing, how the candidate hits, and whether the loop closure verification is finally passed).

[0158] field name meaning RecordID Determine the primary key of the record ImplantID Implant Parts Marking AbutmentID Base Part Identification ResultType EXPLICIT / INFERRED / UNDETERMINED TemplateID Template Number (Fixed Paragraph Sequence Template) ChainSegments Chain details: pos, rel, u, v, dir, source, sim (sim can be empty relative to original) Breakpoint Breakpoint information: anchor, missingRel, pos (can be empty if explicitly closed) CandidateEdges Candidate edge list: u, rel, v, pos, sim (used for review and statistical analysis of Hit@K / MRR) ConfigVersion Dictionary and hyperparameter version numbers (to ensure reproducibility of the experiment) Timestamp Judgment Time

[0159] The experiment uses part pairs as the sample unit (ID_I, ID_A), and the label y∈{1,0} represents compatibility / incompatibility. The breakpoint completion related indicators are only counted on a subset of the original retrieval breakpoints to avoid mixing in samples that have already been explicitly completed, which would lead to an artificially high closure rate.

[0160] subset Number of sample pairs Compatible (1) Incompatible (0) Used for breakpoint completion evaluation Train 18000 14400 3600 5200 Val 4000 3200 800 1120 Test 6000 4950 1050 1740

[0161] Indicator Explanation:

[0162] Precision: The proportion of samples predicted to be compatible that are actually compatible;

[0163] Recall: The proportion of truly compatible samples that are retrieved as compatible by the method;

[0164] F1: A trade-off between Precision and Recall;

[0165] Accuracy: The overall percentage of correct predictions;

[0166] ClosureRate (Loop Closure Success Rate): Statistically calculated only in samples from the original retrieval breakpoint, representing the proportion of samples that can be closed after breakpoint completion and re-examination;

[0167] Hit@K / MRR: Only evaluate the quality of candidate edges with missing relations. Hit@K checks whether the correct edge is in the top K, while MRR checks the average ranking quality of the correct edge.

[0168] The comparison settings emphasize the differences in capabilities from weakest to strongest:

[0169] A: Rule / field matching (baseline) does not follow the template chain, does not generate breakpoints, and does not perform completion or review;

[0170] B: Template-only retrieval (without completion), can output explicit closed chains and breakpoints, but has no recovery capability after a breakpoint;

[0171] C: Complete method, template retrieval + Node2Vec breakpoint completion + enhanced review loop;

[0172] method Precision Recall F1 Accuracy ClosureRate Hit@10 MRR A rule / field matching 0.842 0.711 0.771 0.790 — — — B Template-only search 0.914 0.458 0.610 0.704 0.000 — — C complete method 0.889 0.914 0.901 0.835 0.741 0.781 0.492

[0173] Interpretation of Results (Explanation of Differences by Mechanism):

[0174] B has a high precision but a low recall because the template chain is strict and breaks if an edge is missing. It is conservative but misses a lot of cases.

[0175] C significantly improves Recall and F1, because breakpoint completion can bring the originally broken samples back to a closed state. At the same time, due to the closed-loop constraint of re-examination, Precision remains at a high level without significant drift.

[0176] A has a decent recall but low precision. The typical reason is that field rules are prone to mismatch or omission of key constraints in synonymous encoding / generational mapping / real data with missing fields, and lack chain evidence and segmentation constraints.

[0177] Ablation only modifies one key component, keeping the rest consistent with the full method, to answer which step brought the gain:

[0178] C1: Removing the template consistency constraint during the walk causes the walk to no longer follow the order of missing segments + remaining segments, resulting in more noisy learned context;

[0179] C2: Removing the conditional embedding of relation types and only doing node embedding will mix the completion tasks of different missing relations together, making the candidates more likely to go astray.

[0180] C3: Remove the re-verification and output the inference compatibility directly after patching the edges, which is equivalent to replacing the closed-loop evidence with similarity;

[0181] variants Precision Recall F1 Accuracy ClosureRate Hit@10 MRR C complete method 0.889 0.914 0.901 0.835 0.741 0.781 0.492 C1 w / o template consistent walk 0.861 0.882 0.871 0.804 0.611 0.652 0.401 C2 w / o relational conditional embedding 0.874 0.896 0.885 0.816 0.664 0.701 0.438 C3 w / o re-inspection and verification 0.781 0.945 0.856 0.742 0.741 0.781 0.492

[0182] Ablation conclusion:

[0183] C1 / C2 mainly affect the quality of candidate edges (Hit@10, MRR) and the closure rate (ClosureRate), indicating that walks aligned with template segment order and relation conditionalization are the core factors that enable breakpoint completion to function stably.

[0184] C3 has a high recall but a significant decline in precision, indicating that the closed loop of the re-examination is the key line of defense against misjudgment. Without verifying the closed loop, the edge filling is more like guessing, and it is easier to push incompatible samples to be compatible.

[0185] This table only counts breakpoint samples, focusing on whether the candidate edge of the missing relation can hit the real edge;

[0186] method Hit@1 Hit@3 Hit@10 MRR General Node2Vec (unconstrained) 0.214 0.372 0.603 0.341 Constraint walk + conditional embedding (this method) 0.352 0.541 0.781 0.492

[0187] Summary Table

[0188] This table represents the final delivery metrics: it shows the business the quantity of each of the three types of outputs, and the algorithm the number of breakpoints, loop closures, and loop closure rates. It also places the overall Precision / Recall / F1 in the same table for easy reporting.

[0189] Total number of tests Explicit compatibility Inference Compatibility Undetermined Original number of search breakpoints Recheck Closure Number ClosureRate Precision Recall F1 2480 2610 910 3520 2610 0.741 0.889 0.914 0.901

[0190] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for determining graft connectivity compatibility based on Node2Vec and compatibility knowledge graph retrieval, characterized in that, include: S1. Receive the specified implant identifier and the specified abutment identifier, and generate a compatibility knowledge graph based on the inventory parts master data and manufacturer connection information. S2. Determine the connection chain template based on the compatibility knowledge graph. The connection chain template defines the connectability determination constraints in the order of connection form matching, platform diameter matching, screw specification matching, and series generational mapping. Search the connection chain from the specified implant node to the specified abutment node in the compatibility knowledge graph according to the connection chain template to obtain the search status. The search status is either a closed connection chain or a breakpoint description. The breakpoint description includes anchor nodes and missing relationship types. S3. Based on the breakpoint description, perform Node2Vec representation learning on anchor nodes in the compatibility knowledge graph to obtain the node representation set. Determine the target node set based on the missing relation type. Calculate the representation similarity between anchor nodes and target nodes. If the representation similarity meets the similarity threshold, generate a candidate compatibility relation list. The candidate compatibility relation list includes candidate relations and candidate relation location identifiers. S4. Based on the candidate compatibility relationship list, write the candidate relationships into the compatibility knowledge graph according to the candidate relationship position identifier to form an enhanced knowledge graph, and retrieve the verification connection chain in the enhanced knowledge graph according to the connection chain template; S5. Determine the compatibility judgment result based on the retrieval status and the verification connection chain. When the retrieval status is a closed connection chain, determine the explicit compatibility judgment result and associate it with the closed connection chain. When the retrieval status is a breakpoint description and the verification connection chain is a closed connection chain, determine the inferred compatibility judgment result and associate it with the verification connection chain. When the verification connection chain is not a closed connection chain, determine the result that cannot be judged and associate it with the breakpoint description. S6. Generate a connection compatibility determination record based on the compatibility determination result. The connection compatibility determination record contains the compatibility determination result and the closed connection chain, verification connection chain or breakpoint description associated with the compatibility determination result.

2. The method for determining the compatibility of implantation remediation connections based on Node2Vec and compatibility knowledge graph retrieval as described in claim 1, characterized in that, S1 specifically refers to: Receive a specified implant identifier and a specified abutment identifier, locate the implant part record and abutment part record in the inventory parts master data, and write the specified implant identifier and the specified abutment identifier into the corresponding node primary key respectively; Extract the connection type, platform diameter, screw specification, and series generation fields from the inventory parts master data and manufacturer connection data, and map the field values ​​to preset enumeration values ​​to form a set of specification values ​​that can be used for relationship construction. The implant component record and the abutment component record are used to generate implant node and abutment node, the specification value set is used to generate specification node, and the association relationship between component node and specification node is established to solidify the interface specification of the component. Based on the manufacturer's connection data and the aforementioned relationships, connection form matching relationships, platform diameter matching relationships, screw specification matching relationships, and generational mapping relationships are generated. These four types of relationships are written into the relationship type set, and the starting node, ending node, and relationship type are recorded in the edge set. This results in a compatibility knowledge graph containing the node set, edge set, and relationship type set. Node encoding tables and relationship type encoding tables are also generated for one-hot encoding input in subsequent representation learning.

3. The method for determining graft connectivity compatibility based on Node2Vec and compatibility knowledge graph retrieval as described in claim 1, characterized in that, S2 specifically refers to: Based on the set of relationship types in the compatibility knowledge graph, four types of relationship types are identified: connection form matching relationship, platform diameter matching relationship, screw specification matching relationship, and series generation mapping relationship. The relationship type sequence is formed in the order of connection form matching relationship, platform diameter matching relationship, screw specification matching relationship, and series generation mapping relationship. The relation type sequence is defined as the constraint segment sequence of the connection chain template, and each constraint segment is bound with the allowed retrieval relation direction and segment order position identifier to obtain the connection chain template with single path constraint. Using a specified implant node as the starting retrieval node, the successor node that satisfies the corresponding relationship type is retrieved in the compatibility knowledge graph according to the first constraint segment of the connection chain template, and the successor node is determined as the first intermediate landing point. Using the first intermediate landing point as the current node, continue to search and update the intermediate landing point according to the next constraint segment of the connection chain template, so that the search process forms an intermediate landing point corresponding to the segment order position identifier in each constraint segment. When all constraint segments of the connection chain template are completed and the final intermediate landing point is the specified base node, the relation instances corresponding to each constraint segment are concatenated according to the constraint segment sequence to generate a closed connection chain, and the closed connection chain is determined as the retrieval state. When a successor node cannot be found based on the next constraint segment during the retrieval process, the intermediate point of the last successful update is determined as the anchor node, the relation type corresponding to the next constraint segment is determined as the missing relation type, a breakpoint description containing the anchor node and the missing relation type is generated, and the breakpoint description is determined as the retrieval status.

4. The method for determining the compatibility of implantation remediation connections based on Node2Vec and compatibility knowledge graph retrieval as described in claim 3, characterized in that, In the process of continuing to search and update intermediate landing points according to the next constraint segment of the connection chain template, with the first intermediate landing point as the current node, the template constraint reachability score of each candidate successor node obtained from the next constraint segment search is calculated to the specified base node. Based on the template constraint reachability score, an intermediate landing point for updating is selected and determined from the candidate successor nodes. The template constraint reachability score is calculated by a constraint function, which is specifically: ； in, Indicates a candidate successor node Currently in the th When constraining segments, along the connection chain template The remaining constraint segment reaches the specified base node. Template constraint path count, This represents any node in the set of successor nodes obtained from the current constraint segment retrieval. This indicates the sequence number of the constraint segment that has been completed. Indicates a link template The sequence number variable of the middle constraint segment. This indicates the relationship between nodes generated from the node encoding table and nodes. The corresponding one-hot vector, This indicates the node generated by the node encoding table and the specified base station node. The corresponding one-hot vector, This represents the vector transpose operation. Indicates a link template The Middle Segment constraint segment relationship type, Indicates a link template The Middle The direction of the relationship between segment constraints and segment bindings. Indicates the relation type is And the direction is The relational adjacency matrix uses the node indices of the node encoding table as coordinates for its rows and columns. Matrix elements are set to one when an edge satisfying the relation type and direction constraints exists, pointing from a row node to a column node; otherwise, they are set to zero. Indicates by from arrive Perform matrix chain multiplication in the order described For connecting chain templates The total number of constraint segments included.

5. The method for determining the compatibility of implantation remediation connections based on Node2Vec and compatibility knowledge graph retrieval as described in claim 1, characterized in that, S3 specifically refers to: Read the breakpoint description to obtain the anchor node and the missing relationship type, and determine the segment position identifier corresponding to the missing relationship type according to the connection chain template; Using the anchor node as the starting point of the walk, the first jump edge is selected in the compatibility knowledge graph according to the missing relation type constraint, and the subsequent jump edges are selected according to the subsequent relation type sequence constraint of the connection chain template, generating a walk sequence that is consistent with the order of the connection chain template; Training samples are constructed based on the walking sequence. The central node and the context node are determined as node encoding features, and the relationship type from the central node to the context node is determined as the relationship type encoding feature. The node encoding features are input into the node encoding neurons and represented using one-hot encoding. The relation type encoding features are input into the relation type encoding neurons and represented using one-hot encoding. The value of the relation type encoding features is determined by the missing relation type. The node encoding features are mapped to node representations through the node embedding layer, the relation type encoding features are mapped to relation type representations through the relation type embedding layer, and the node representations and relation type representations are added together in the fusion layer to obtain the center node representation with missing relation type conditionalization. The conditional center node representation is predicted by the context node prediction layer. Negative sampling is used to train and update the parameters of the node embedding layer, relation type embedding layer and context node prediction layer to obtain a set of node representations including anchor nodes. The target node set is determined based on the relationship pointing range of the missing relationship type in the compatibility knowledge graph. The representation similarity between the anchor node and each target node in the target node set is calculated. A candidate compatible relationship list is generated when the representation similarity meets the similarity threshold. The candidate relationship in the candidate compatible relationship list consists of the anchor node, the missing relationship type, and the target node. The position identifier of the candidate relationship is determined by the segment position identifier.

6. The method for determining the compatibility of implantation remediation connections based on Node2Vec and compatibility knowledge graph retrieval as described in claim 5, characterized in that, Negative sampling is used to train and update the parameters of the node embedding layer, relation type embedding layer, and context node prediction layer. This includes: calculating the log loss value for each training sample based on the missing relation type conditional center node representation, the context node prediction layer output, and the negative sample nodes obtained from negative sampling; and performing parameter updates based on the log loss value. The log loss value is calculated by a loss function, which is specifically: ； in, This represents the logarithmic loss value for a single training sample. Represents the sigmoid activation function. This represents the central node output by the node embedding layer. The node represents a vector. This indicates the missing relation type output by the relation type embedding layer. Relational types represent vectors. This indicates the transpose operation. Represents the context node The output vector in the context node prediction layer. This indicates the number of nodes sampled according to their frequency distribution. One negative sample node Represents negative sample nodes The output vector in the context node prediction layer. This indicates the number of negative samples.

7. The method for determining the compatibility of implantation remediation connections based on Node2Vec and compatibility knowledge graph retrieval as described in claim 1, characterized in that, S4 specifically refers to: Read the candidate compatibility relationship list, group the candidate relationships according to their position identifiers, and obtain the candidate relationship subsets corresponding to each constraint segment of the connection chain template; For each candidate relation, the anchor node, missing relation type, and target node are parsed, and the missing relation type is determined as the write relation type; Based on the candidate relationship position identifier, the write relationship type is bound to the segment order position identifier of the corresponding constraint segment in the connection chain template, forming a one-to-one correspondence between the candidate relationship and the constraint segment; Write candidate edges with anchor nodes as start nodes, target nodes as end nodes, and the type of relation to be written into the edge set of the compatibility knowledge graph, while keeping the original edge set in the compatibility knowledge graph from being replaced, to obtain the enhanced knowledge graph. Based on the connection chain template, a structured retrieval is performed in the enhanced knowledge graph with a specified implant node as the starting retrieval node. During the retrieval process, only relation types and their corresponding candidate edges that are consistent with the segment order position identifier of each constraint segment are allowed to be used for each constraint segment. During the retrieval process, intermediate landing points are updated segment by segment and relational instances are connected. When all constraint segments of the connection chain template are completed and the final intermediate landing point is the specified base node, a verification connection chain is generated. The source of the relation instances used in each constraint segment in the verification connection chain is the original edge or candidate edge of the compatibility knowledge graph, and the output of the verification connection chain is used for compatibility determination.

8. The method for determining the compatibility of implantation remediation connections based on Node2Vec and compatibility knowledge graph retrieval as described in claim 1, characterized in that, S5 specifically refers to: Read the search status and determine whether the search status is a closed link chain. At the same time, read the verification link chain and determine whether the verification link chain is a closed link chain in the order of the link chain template. When the retrieval status is a closed connection chain, the compatibility determination result is determined as an explicit compatibility determination result, and the closed connection chain is used as the associated connection chain of the compatibility determination result; When the retrieval status is a breakpoint description and the verification connection chain is a closed connection chain, the compatibility determination result is determined as the inferred compatibility determination result, and the verification connection chain is used as the associated connection chain of the compatibility determination result; When verifying that the connection chain is not a closed connection chain, the compatibility determination result is determined as an undeterminable result, and the breakpoint description is used as the associated breakpoint description of the compatibility determination result.

9. The method for determining the compatibility of implantation remediation connectivity based on Node2Vec and compatibility knowledge graph retrieval as described in claim 8, characterized in that, Reading the verification link and determining whether it is a closed link in the order of the link template includes: calculating a closure determination value for the verification link and the link template, and using the value of the closure determination value to determine whether the verification link is a closed link in the order of the link template. The closure determination value is calculated by a determination function, which is specifically: ； in, This represents the closure criterion value, which takes the value of... Indicates verification of the connection chain To connect the template A sequentially closed chain of connections takes the value of Indicates verification of the connection chain Not a closed link chain This indicates an indicator function that takes the value when the condition within the parentheses is true. Otherwise, the value is , Indicates verification of the connection chain The node sequence, Indicates the starting node. Indicates the final node, Indicates the specified implant node. Indicates the specified base station node. Indicates verification of the connection chain Number of relation instances Indicates a link template The number of constraint segments, Indicates verification of the connection chain The Middle The relation type of each relation instance, Indicates a link template The Middle Segment constraint segment binding relationship type, Indicates verification of the connection chain The Middle Directional encoding of relation instances, Indicates a link template The Middle Segment constraint segment binding relationship direction encoding, Indicates by from arrive Perform a series of multiplications.

10. The method for determining the compatibility of planting remediation connections based on Node2Vec and compatibility knowledge graph retrieval as described in claim 1, characterized in that, Step S6 is as follows: Read the compatibility determination result, extract the compatibility determination result type and the closed connection chain, verification connection chain or breakpoint description associated with the compatibility determination result; Write the specified implant node, specified abutment node, connection chain template, and compatibility determination result type into the main field of the connection compatibility determination record; When the closed connection chain or verification connection chain associated with the compatibility determination result exists, the relation type, start node, end node, segment order position identifier, and relation instance source as the original edge or candidate edge of the compatibility knowledge graph corresponding to each constraint segment in the connection chain are written into the chain field of the connection compatibility determination record. When the breakpoint description associated with the compatibility determination result exists, the anchor node, the missing relationship type, and the segment position identifier corresponding to the missing relationship type are written into the breakpoint field of the connection compatibility determination record to generate the connection compatibility determination record.

Images