Injection molding process fault diagnosis system based on knowledge graph

Abstract

The invention discloses an injection molding process fault diagnosis system based on a knowledge graph, and relates to the technical field of injection molding process fault diagnosis. The injection molding process data comprises novel fault case data, maintenance and solution records, expert experience and feedback and new equipment specifications, data preprocessing and structuring are carried out, and a multi-source dynamic data stream is output; and the knowledge increment extraction verification module is used for executing entity identification and relation extraction based on the multi-source dynamic data flow. Multi-dimensional data such as novel fault cases and expert experience are captured in real time through the multi-source data dynamic acquisition module, and the problem that a static knowledge graph cannot be brought into new changes is solved; the knowledge increment extraction verification and self-adaptive updating module can dynamically expand the ontology, incrementally update the atlas and guarantee the data consistency, ensure that the atlas can incorporate fault knowledge related to new materials and new equipment in real time, and expand the fault coverage range.

Description

Technical Field

[0001] This invention relates to the field of injection molding process fault diagnosis technology, specifically to an injection molding process fault diagnosis system based on knowledge graphs. Background Technology

[0002] Injection molding, a widely used molding technology in modern manufacturing, involves the interaction of multiple factors such as material properties, equipment parameters, and environmental conditions, resulting in complex and diverse causes of failures. Knowledge graphs, with their structured representation capabilities of entities, relationships, and rules, are widely used in the field of injection molding fault diagnosis. By integrating knowledge nodes such as abnormal equipment parameters, material defects, and process parameter deviations, a network linking faults and their causes is constructed, providing knowledge support for quickly locating the root cause of the fault.

[0003] However, most existing injection molding process knowledge graphs are statically constructed, meaning that the entity types, relationship rules, and fault solutions they contain remain fixed after construction. With the continuous advancement of new material applications, equipment upgrades, and process optimization, new fault modes constantly emerge during injection molding production, and existing fault solutions also need to be dynamically updated. Static knowledge graphs struggle to incorporate these new changes in real time, leading to a gradual narrowing of their fault coverage and a continuous decline in diagnostic accuracy, failing to meet the dynamic needs of fault diagnosis in actual production. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a knowledge graph-based injection molding process fault diagnosis system. This system solves the problem that existing injection molding process knowledge graphs are mostly statically constructed, and the entity types, relationship rules, and fault solutions they contain are fixed after construction, making it difficult to incorporate new changes in real time.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a knowledge graph-based injection molding process fault diagnosis system, the system comprising: The multi-source dynamic data acquisition module is used to collect injection molding process data, which includes new fault case data, maintenance and solution records, expert experience and feedback, and new equipment specifications. The module performs data preprocessing and structuring, and outputs a multi-source dynamic data stream. The knowledge increment extraction and verification module, based on the multi-source dynamic data stream, performs entity recognition and relation extraction, performs rule learning and pattern discovery, detects conflicts between newly added knowledge and existing knowledge graphs, completes preliminary knowledge verification, and obtains incremental knowledge units to be integrated. The knowledge graph adaptive update module receives the incremental knowledge units to be integrated, performs ontology mapping and alignment operations, calls the incremental graph update algorithm, adds, deletes, and modifies nodes in the injection molding process knowledge graph, records graph version information, and generates an updated injection molding process knowledge graph. The closed-loop feedback and reasoning module, based on the updated injection molding process knowledge graph and combined with real-time fault symptoms, performs reasoning to generate diagnostic results. It receives expert correction and confirmation information on the diagnostic results through a human-computer interaction interface, and transmits the correction and confirmation information as expert experience and feedback to the multi-source data dynamic acquisition module.

[0006] Furthermore, the multi-source dynamic data stream includes structured fault logs, unstructured maintenance report texts, expert feedback rule sets, and standardized process parameter tables; the incremental knowledge units to be integrated include new fault entities, new causal triples, new diagnostic rules, and extended ontology concepts; the updated injection molding process knowledge graph includes updated equipment component nodes, process parameter edges, fault phenomenon attributes, and solution paths; and the correction confirmation information includes diagnostic result correctness labels, knowledge graph node correction suggestions, and newly added experience rule texts.

[0007] Furthermore, the multi-source data dynamic acquisition module includes: The real-time data capture submodule is used to access the equipment operation log system and alarm record database, capture new fault case data in real time, filter fault records not covered by the knowledge graph, and generate fault log sequences. The semi-structured information parsing submodule calls the fault log sequence to parse the operation steps, replaced parts and adjusted parameters in the maintenance and solution records, extracts key information fields, and generates structured maintenance records; The expert knowledge acquisition submodule receives expert experience and feedback through a human-computer interaction interface, converts the correction opinions or new rules described in natural language into a predefined format, and generates expert knowledge data. The standardized document processing submodule is used to process technical manuals and process specification documents for new equipment and materials, extract key performance parameters and operational constraints, and generate standardized equipment specification data.

[0008] Furthermore, the knowledge increment extraction verification module includes: The hybrid entity relation extraction submodule, based on the multi-source dynamic data stream, uses natural language processing technology to perform named entity recognition on unstructured text, identifies entities such as process parameters, fault phenomena, and solutions, and uses a relation extraction model to extract the associations between entities and generate candidate knowledge triples. The association rule mining submodule analyzes sensor readings and fault records in structured data, uses machine learning algorithms to discover potential causal relationships and fault modes between data, and generates candidate diagnostic rules. The knowledge conflict detection submodule compares the candidate knowledge triples and candidate diagnostic rules with the injection molding process knowledge graph to detect logical conflicts, factual contradictions, and data redundancy, and marks conflicting knowledge points. The knowledge rationality verification submodule calls the preset physical constraint rules and domain knowledge model to conduct a preliminary rationality assessment of candidate knowledge that does not conflict, selects the knowledge that passes the verification, and obtains the incremental knowledge units to be integrated.

[0009] Furthermore, the hybrid entity relationship extraction submodule uses a pre-trained language model based on a transformer to process the maintenance report text, identify the fault cause entity and solution entity contained therein, and calculate the semantic association strength between the two. When the semantic association strength exceeds a preset association threshold, a fault-cause-solution relationship triplet is generated.

[0010] Furthermore, the knowledge graph adaptive update module includes: The ontology dynamic extension submodule analyzes whether the new entity or new relationship exceeds the definition range of the current ontology based on the incremental knowledge unit to be integrated. If it does, the ontology structure is dynamically extended, new concept classes or attribute definitions are added, and the extended ontology is generated. The incremental update execution submodule calls the extended ontology to map the entities and relations in the incremental knowledge units to be integrated to graph nodes and edges. It uses incremental graph database transaction operations to add nodes, modify attributes, and delete edges to ensure the atomicity of the update process. The version control and rollback submodule records a snapshot of the knowledge graph's state and a change log before and after each update operation. It assigns a unique version number to each successful update and supports rolling back the knowledge graph to a specified historical state based on the version number.

[0011] Furthermore, the incremental update execution submodule adopts a write-ahead logging mechanism. Before actually modifying the map data file, the update operation is recorded in the log file to ensure that data consistency can be restored by replaying the log in the event of a system failure, thus ensuring the reliability of the update.

[0012] Furthermore, the closed-loop feedback and inference module includes: The multi-hop inference diagnosis submodule receives real-time fault symptoms as input, performs multi-hop graph traversal in the updated injection molding process knowledge graph, searches for the shortest or highest confidence path from the fault phenomenon node to the root cause node, and generates a preliminary diagnosis result containing the inference path. The diagnostic results presentation submodule visualizes the preliminary diagnostic results on the human-computer interaction interface, clearly presenting the fault phenomenon, the possible causes, and the corresponding solution suggestions. The expert feedback quantification submodule provides experts with scoring of diagnostic results, correction text boxes, and interfaces for adding new knowledge entries. It transforms unstructured feedback input by experts into structured correction and confirmation information, which includes instructions for adjusting the confidence level of specific triples in the graph.

[0013] Furthermore, the system also includes a knowledge graph health assessment module; The knowledge graph health assessment module periodically evaluates the quality of the injection molding process knowledge graph, calculates knowledge coverage, knowledge redundancy, knowledge consistency and expert feedback adoption rate, and generates a health assessment report. The knowledge coverage rate represents the proportion of fault cases that the current knowledge graph can explain to the total fault case library; the expert feedback adoption rate represents the proportion of expert feedback that has been integrated into incremental knowledge units to the total number of expert feedback.

[0014] Furthermore, the knowledge graph health assessment module includes: The coverage calculation submodule counts the number of new fault case data collected by the multi-source data dynamic acquisition module within a certain period of time that can be successfully diagnosed by the closed-loop feedback and reasoning module, calculates the ratio of this number to the total number of new cases, and obtains the knowledge coverage rate. The consistency verification submodule randomly extracts knowledge paths from the knowledge graph, applies predefined logical rules to perform reasoning, detects whether there are contradictory reasoning results, calculates the proportion of contradictory paths to the total sampled paths, and obtains the knowledge inconsistency index. The feedback value analysis submodule tracks the correction confirmation information generated by the expert feedback quantification submodule, calculates the proportion of the information that is adopted by the knowledge increment extraction and verification module and finally updated in the graph, and obtains the expert feedback adoption rate. Beneficial effects

[0015] This invention solves the problem of static knowledge graphs being unable to incorporate new changes by using a multi-source data dynamic acquisition module to capture new fault cases and expert experience in real time. The knowledge increment extraction verification and adaptive update module dynamically expands the ontology, incrementally updates the graph, and ensures data consistency, guaranteeing that the graph can incorporate fault knowledge related to new materials and equipment in real time, thus expanding the fault coverage. The closed-loop feedback and reasoning module combines expert feedback to optimize diagnostic results and improve diagnostic accuracy. The knowledge graph health assessment module quantitatively monitors graph quality, providing direction for system optimization. In summary, this system achieves dynamic iteration of the knowledge graph, effectively improving the timeliness and accuracy of fault diagnosis and meeting the dynamic needs of actual production. Attached Figure Description

[0016] Figure 1 This is a system structure diagram of the present invention. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] Please see Figure 1 This invention provides a knowledge graph-based injection molding process fault diagnosis system. The system includes a multi-source data dynamic acquisition module for acquiring injection molding process data, which includes new fault case data, maintenance and solution records, expert experience and feedback, and new equipment specifications. The system performs data preprocessing and structuring, and outputs a multi-source dynamic data stream. The knowledge increment extraction and verification module, based on multi-source dynamic data stream, performs entity recognition and relation extraction, rule learning and pattern discovery, detects conflicts between new knowledge and existing knowledge graphs, completes preliminary knowledge verification, and obtains incremental knowledge units to be integrated. The knowledge graph adaptive update module receives incremental knowledge units to be integrated, performs ontology mapping and alignment operations, calls the incremental graph update algorithm, adds, deletes, and modifies nodes in the injection molding process knowledge graph, records graph version information, and generates an updated injection molding process knowledge graph. The closed-loop feedback and reasoning module, based on the updated injection molding process knowledge graph and combined with real-time fault symptoms, infers and generates diagnostic results. It receives expert correction and confirmation information on the diagnostic results through a human-computer interaction interface, and transmits the correction and confirmation information as expert experience and feedback to the multi-source data dynamic acquisition module.

[0019] Specifically, the system addresses the issue of static knowledge graphs failing to incorporate new changes in the background technology through a multi-source data dynamic acquisition module. This module uses an industrial Ethernet connection to the equipment control system in the injection molding workshop, collecting 16 types of process parameters in real time, such as injection speed and holding pressure, at a sampling frequency of 1kHz. Simultaneously, it receives paper maintenance reports via a document scanning interface, converting them into text data after OCR recognition. For example, when a new bio-based material is introduced into the workshop, the system automatically captures parameters such as the material's molding temperature range and shear rate, generating standardized equipment specification data.

[0020] In the knowledge incremental extraction and verification module, entity recognition uses the BERT-BiLSTM-CRF model to identify the entities "new material degradation causes product yellowing" and "product yellowing" as the causes of the failure for text like "new material degradation causes product yellowing". Relation extraction uses an attention mechanism to calculate entity association, with the formula as follows: ,in, The current entity vector is either a word vector or an entity vector generated by a pre-trained model. These are the attention weights, the parameters obtained during model training. The total number of entities is an integer count. A relation triple is generated when the association degree exceeds 0.65.

[0021] The knowledge graph adaptive update module uses the graph database Neo4j. The incremental update algorithm is based on the RDF incremental loading mechanism. It follows a process of "mapping first, then verifying, and then updating". For example, it maps "bio-based materials" to a subclass of the "materials" class and adds the "degradation temperature" attribute.

[0022] The closed-loop feedback and reasoning module employs a bidirectional breadth-first search for multi-hop reasoning. Starting from the real-time fault symptom "surface cracking of the product," it searches for the path "material degradation → molecular chain breakage → surface cracking," generating a diagnostic result. Experts annotate the accuracy of this result via an interface, and the correction information is transmitted back to the multi-source data acquisition module in real time. Through this implementation method, the system can dynamically incorporate fault knowledge related to new materials and equipment, improving fault coverage by more than 60% compared to a static system.

[0023] In this embodiment, the multi-source dynamic data stream includes structured fault logs, unstructured maintenance report texts, expert feedback rule sets, and standardized process parameter tables; the incremental knowledge units to be integrated include new fault entities, new causal triples, new diagnostic rules, and extended ontology concepts; the updated injection molding process knowledge graph includes updated equipment component nodes, process parameter edges, fault phenomenon attributes, and solution paths; and the correction confirmation information includes diagnostic result correctness labels, knowledge graph node correction suggestions, and newly added experience rule texts.

[0024] Specifically, in the multi-source dynamic data stream, the structured fault log adopts JSON format, containing fields such as "Equipment No.: M202305, Fault Time: 2024-03-12 08:45, Fault Code: E109, Parameter Sequence: [Temperature: 230℃, Pressure: 110bar,...]"; the unstructured maintenance report text is a Word format maintenance record, such as "On March 12, 2024, the M202305 equipment was inspected and the check ring was found to be worn. After replacement, the melting speed was adjusted from 70rpm to 65rpm, and the fault was resolved." In the incremental knowledge units to be integrated, new fault entities such as "bio-based material hydrolysis" are generated by analyzing the description of "bubbles appearing in products under humid conditions" in the maintenance report; new causal triples such as "check ring wear → injection pressure fluctuation" are obtained by association rule mining; new diagnostic rules adopt the "IF-THEN" form, for example, "IF ambient humidity > 60% AND material is bio-based THEN, preheating and drying for 2 hours is recommended"; extended ontology concepts such as "intelligent check ring" are added as a subclass of "equipment component". In the updated injection molding process knowledge graph, the equipment component node adds "intelligent check ring", whose attributes include "wear warning threshold"; the weight of the process parameter edge "melt speed - product density" is adjusted according to the new data; the fault phenomenon attribute "bubbles" adds the "distribution location" dimension; the solution path is refined into a step chain of "replacing the check ring → adjusting the melt speed → detecting the product density". In the revised confirmation information, the correctness label for the diagnostic result is "Correct / Incorrect / Partially Correct"; the knowledge graph node revision suggestions are such as "revise the 'fluctuation range' attribute of 'injection pressure fluctuation' from ±10 bar to ±8 bar"; the newly added empirical rule text, after being structured, is "When using bio-based materials, the ambient temperature should be controlled at 25±2℃". This implementation method makes the form and flow of knowledge data clearer, providing accurate input for dynamic updates of the graph.

[0025] In this embodiment, the multi-source data dynamic acquisition module includes: a real-time data capture submodule, used to access the equipment operation log system and alarm record database, capture new fault case data in real time, filter fault records not covered by the knowledge graph, and generate fault log sequences; a semi-structured information parsing submodule, which calls the fault log sequences, parses the operation steps, replacement parts, and adjustment parameters in the maintenance and solution records, extracts key information fields, and generates structured maintenance records; an expert knowledge acquisition submodule, which receives expert experience and feedback through a human-computer interaction interface, converts the correction opinions or new rules described in natural language into predefined formats, and generates expert knowledge data; and a standardized document processing submodule, used to process technical manuals and process specification documents for new equipment and new materials, extract key performance parameters and operating constraints, and generate standardized equipment specification data.

[0026] Specifically, the real-time data capture submodule connects to the equipment PLC's database via a JDBC interface and sets up a fault feature matching algorithm. For example, when a "sudden drop in injection pressure > 30% and duration > 5s" is detected and this feature does not exist in the knowledge graph, it is determined to be a new type of fault, and the parameter sequence of the first 30s and the last 30s is automatically extracted to generate a fault log. The semi-structured information parsing submodule uses rule-based template matching. For example, from the maintenance record "Replaced Φ38mm screw, adjusted back pressure to 9 bar, passed after 5 tests", fields such as "Replaced part: Φ38mm screw", "Adjusted parameters: back pressure = 9 bar", and "Verification result: Pass" are extracted to generate a structured maintenance record. The expert knowledge acquisition submodule designs an interactive form. After the expert selects the fault type "product warping", they enter "When using glass fiber reinforced materials, the holding time should be extended by 15%" in the text box. The system automatically converts it into the rule format "Material = Glass fiber reinforcement → Holding time = Original time × 1.15". The standardized document processing submodule uses the PDFMiner tool to parse new equipment manuals. For example, it extracts parameters such as "maximum injection volume: 500g," "screw diameter: 45mm," and "injection rate: 200mm / s" from the "HTF300W injection molding machine technical parameters," generating standardized data in the structure of "equipment model-parameter name-parameter value-unit." This implementation method improves the efficiency of multi-source data acquisition and structuring by 50%, ensuring that new faults and process knowledge are captured in a timely manner.

[0027] In this embodiment, the knowledge increment extraction and verification module includes: a hybrid entity relationship extraction submodule, which uses natural language processing technology to perform named entity recognition on unstructured text based on multi-source dynamic data streams, identifies entities such as process parameters, fault phenomena, and solutions, and uses a relationship extraction model to extract the associations between entities to generate candidate knowledge triples; an association rule mining submodule, which analyzes sensor readings and fault records in structured data, uses machine learning algorithms to discover potential causal relationships and fault modes between data, and generates candidate diagnostic rules; a knowledge conflict detection submodule, which compares the candidate knowledge triples and candidate diagnostic rules with the injection molding process knowledge graph to detect logical conflicts, factual contradictions, and data redundancy, and marks conflicting knowledge points; and a knowledge rationality verification submodule, which calls preset physical constraint rules and domain knowledge models to perform a preliminary rationality assessment on candidate knowledge that does not conflict, filters out knowledge that passes verification, and obtains incremental knowledge units to be integrated.

[0028] Specifically, the hybrid entity relation extraction submodule first uses the spaCy tool for word segmentation and part-of-speech tagging of unstructured repair report text, and then identifies entities using the BERT model. For example, from the sentence "The mold temperature is too low, causing weld lines on the product, requiring the mold temperature to be increased to 60℃," it identifies the cause of the fault "the mold temperature is too low," the phenomenon of the fault "weld lines on the product," and the solution of "increasing the mold temperature to 60℃." Relation extraction uses the PCNN model to calculate the association probability between entities, generating candidate triples such as "The mold temperature is too low → causing → weld lines on the product" and "Weld lines on the product → require → increasing the mold temperature to 60℃." The association rule mining submodule uses the FP-Growth algorithm to mine frequent itemsets from structured sensor data. The support calculation formula is: , in, This is a set of parameters, such as "melt temperature > 240℃ and injection speed > 60mm / s". The fault type is indicated, such as "product burnt". For simultaneous occurrence and The number of records, integer count, For the total number of records, an integer count is used. Candidate rules are generated when the support is greater than 0.2 and the confidence is greater than 0.7, for example, "melt temperature > 240℃ and injection speed > 60mm / s → product scorching". The knowledge conflict detection submodule compares candidate knowledge with the existing graph using the OWLDL inference engine. For example, if the candidate rule "mold temperature should be > 80℃" contradicts the existing rule "mold temperature ≤ 75℃", it is marked as a conflict. The knowledge rationality verification submodule calls the polymer rheology model to verify whether the relationship between "injection speed and apparent melt viscosity" conforms to the power law equation, filtering out knowledge with reasonable physical meaning. This implementation improves the accuracy of incremental knowledge, and the conflict detection accuracy rate can reach over 95%.

[0029] In this embodiment, the hybrid entity relationship extraction submodule uses a pre-trained language model based on Transformer to process the maintenance report text, identify the fault cause entity and solution entity contained therein, and calculate the semantic association strength between the two. When the semantic association strength exceeds the preset association threshold, a fault-cause-solution relationship triplet is generated.

[0030] Specifically, the hybrid entity relation extraction submodule uses a RoBERTa pre-trained model to perform 12-layer Transformer encoding on the input repair report text, generating context-sensitive word vectors. For example, when processing the text "Due to insufficient drying of raw materials, bubbles are generated in the product, and the drying time needs to be extended to 4 hours," the model identifies the entity "insufficient drying of raw materials" (the cause of the fault) and the entity "extend the drying time to 4 hours" (the solution) through the named entity recognition layer, and outputs the vector representations of both. Cause vector sum The scheme vector is generated by a pre-trained model, and the vector elements are L2 normalized to ensure a uniform range.

[0031] The semantic association strength is calculated using cosine similarity, and the formula is: , in, For vector dot product, , This is the vector magnitude. The preset association threshold is set to 0.7. When the calculated value is... The system generates a triplet of relationships: "Product bubbles - cause - insufficient raw material drying" and "Product bubbles - solution - extend drying time to 4 hours". For text containing multiple causes or solutions, such as "Product dimensional deviations may be due to mold wear or unstable injection pressure, requiring mold parts replacement or pressure sensor calibration", the model calculates the correlation strength between each cause and solution, retaining only combinations exceeding a threshold. This implementation improves the accuracy of identifying the correlation between fault causes and solutions to 90%, while reducing the introduction of irrelevant knowledge.

[0032] In this embodiment, the knowledge graph adaptive update module includes: an ontology dynamic expansion submodule, which analyzes whether new entities or relationships exceed the definition range of the current ontology based on the incremental knowledge units to be integrated; if so, it dynamically expands the ontology structure, adds new concept classes or attribute definitions, and generates an expanded ontology; an incremental update execution submodule, which calls the expanded ontology, maps the entities and relationships in the incremental knowledge units to be integrated to graph nodes and edges, and uses incremental graph database transaction operations to add nodes, modify attributes, or delete edges to ensure the atomicity of the update process; and a version control and backtracking submodule, which records the state snapshot and change log of the knowledge graph before and after each update operation, assigns a unique version number to each successful update, and supports backtracking the knowledge graph to a specified historical state based on the version number.

[0033] Specifically, the ontology dynamic expansion submodule uses concept similarity calculation to compare the newly added entity "magnetorheological injection mold" with the existing "traditional mold" class in the ontology to calculate the attribute difference. ,in, For the new entity attribute set, For the existing class attribute set, in terms of the number of attributes, when The system determines that expansion is needed, adding a new "Magnetorheological Injection Mold" class and adding unique attributes such as "Magnetic Field Strength" and "Response Time". The incremental update execution submodule uses the transaction mechanism of the graph database. For example, for the operation of "adding a magnetorheological mold node and its material adaptation relationship", it first executes BEGINTRANSACTION, then sequentially executes CREATE node, CREATE relationship, and finally COMMIT. If any intermediate step fails, it automatically ROLLBACKs to ensure atomicity. The version control and rollback submodule uses a timestamp + sequence number version naming rule, such as V20240315.001. Each update records the details of node / edge additions, deletions, and modifications, for example, "Added node: Magnetorheological Injection Mold ID: 10086; Modified edge: Material-Adaptation-Mold weight changed from 0.6 to 0.8". When an error is found in an update, the rollback interface is called through the version number to restore the system to a historical state such as V20240314.005. This implementation method enables dynamic expansion and secure updates of the knowledge graph, with the ontology expansion response time controlled within 3 minutes.

[0034] In this embodiment, the incremental update execution submodule adopts a pre-write log mechanism. Before actually modifying the map data file, the update operation is recorded in the log file to ensure that data consistency can be restored by replaying the log in the event of a system failure, thus ensuring the reliability of the update.

[0035] Specifically, the pre-write logging mechanism of the incremental update execution submodule follows the process of "log recording → data update → log commit". For example, when executing the operation "modify the standard value of the 'injection pressure' node from 120 bar to 115 bar", the following is first recorded in the log file: [2024-03-15 14:30:00] OP:UPDATE,NodeID:502,Attr:standard value,OldVal:120bar,NewVal:115bar,Checksum:0x3a7b. The log file uses a circular writing method, generating a checkpoint every 100 log entries. During normal system operation, the log recording is completed before... The system performs in-memory data modifications and marks the changes as "committed" in the log upon successful completion. When the system restarts due to a sudden power outage or other failure, the log recovery module scans the log file, replays data marked as "committed" but not yet written to disk, and rolls back operations not marked as "committed." For example, if the failure occurs after the log entry but before the data modification, the log is replayed to complete the data update; if the failure occurs during data modification, the system rolls back to the state before the modification. This mechanism improves the reliability of update operations to 99.99%, avoiding inconsistencies in the graph data caused by partial updates.

[0036] In this embodiment, the closed-loop feedback and reasoning module includes: a multi-hop reasoning diagnosis submodule, which receives real-time fault symptoms as input, performs multi-hop graph traversal in the updated injection molding process knowledge graph, searches for the shortest or highest confidence path from the fault phenomenon node to the root cause node, and generates a preliminary diagnosis result containing the reasoning path; a diagnosis result presentation submodule, which visualizes the preliminary diagnosis result on the human-computer interaction interface, clearly presenting the fault phenomenon, the inferred possible causes, and the corresponding solution suggestions; and an expert feedback quantification submodule, which provides experts with scoring of the diagnosis result, correction text boxes, and an interface for adding new knowledge, transforming the unstructured feedback input by experts into structured correction confirmation information, which includes confidence adjustment instructions for specific triples in the graph.

[0037] Specifically, the multi-hop inference diagnosis submodule uses an improved Dijkstra algorithm. Starting from the fault phenomenon node, it uses the confidence level of the edges as weights (weight = 1 - confidence level) to search for the path with the lowest weight. For example, for the fault "scratches on the surface of the product", the path "scratches on the surface of the product ← caused by... ← impurities in the mold cavity ← caused by... ← incomplete cleaning" is found in the graph. The total weight of the path is 0.2, which is the sum of the weights of each edge, corresponding to a confidence level of 0.8. A preliminary diagnosis result containing this path is generated. The diagnosis result presentation submodule uses a force-directed graph for visualization. The node size represents the importance of the entity, and the color of the edges represents the confidence level (dark color indicates high confidence). The right-hand list displays "Possible cause: impurities in the mold cavity, confidence level 0.8" and "Solution: Stop the machine and clean the cavity. Operation steps: 1. Turn off the heating...". The expert feedback quantification submodule provides a scoring slider from 1 to 5 points. When an expert gives a score of 4, it automatically converts to "Diagnosis result accuracy: good". When the correction text box is entered as "Impurities in the mold cavity may be due to foreign matter in the raw material", the system extracts the relationship "foreign matter in the raw material → impurities in the mold cavity" and generates a confidence adjustment instruction to "increase the confidence of this relationship from 0.3 to 0.6". This implementation method improves the efficiency of expert feedback in fault diagnosis by 40% and accelerates the iterative optimization of the knowledge graph.

[0038] In this embodiment, the system also includes a knowledge graph health assessment module; the knowledge graph health assessment module periodically assesses the quality of the injection molding process knowledge graph, calculates knowledge coverage, knowledge redundancy, knowledge consistency and expert feedback adoption rate, and generates a health assessment report; the knowledge coverage represents the proportion of fault cases that the current knowledge graph can explain to the total fault case library; the expert feedback adoption rate represents the proportion of the number of expert feedbacks integrated into incremental knowledge units to the total number of expert feedbacks.

[0039] Specifically, the knowledge graph health assessment module runs on a weekly cycle.

[0040] Knowledge coverage is calculated as follows: , in, The number of new fault cases that can be explained by the graph. This represents the total number of newly added fault cases, and is an integer. For example, if 50 new faults are added in a week, and 42 can be explained, the coverage rate is 84%. Knowledge redundancy is calculated by the percentage of repeated triples. For example, if the triple "injection speed too fast → flash on product" is detected and stored repeatedly 3 times, the redundancy is 0.5%. Knowledge consistency is checked through logical reasoning. 1000 knowledge paths are randomly selected and verified using the physical rules of the injection molding process. If 5 paths contradict each other, the consistency rate is 99.5%.

[0041] The expert feedback adoption rate was: , in, The number of expert feedback responses that were adopted. The total number of feedback responses is an integer. For example, if 20 feedback responses were received in a week, and 16 were adopted, the adoption rate is 80%. The health assessment report displays the trends of various indicators in the form of a dashboard, and an alert is triggered when the coverage rate falls below 70%. This implementation method enables quantitative monitoring of the knowledge graph quality, providing a clear direction for system optimization.

[0042] In this embodiment, the knowledge graph health assessment module includes: a coverage calculation submodule, which counts the number of new fault case data collected by the multi-source data dynamic acquisition module within a certain period of time that can be successfully diagnosed by the closed-loop feedback and reasoning module, and calculates the ratio of this number to the total number of new cases to obtain the knowledge coverage rate; a consistency verification submodule, which randomly extracts knowledge paths in the knowledge graph, applies predefined logical rules for reasoning, detects whether there are contradictory reasoning results, calculates the proportion of contradictory paths to the total sampled paths, and obtains the knowledge inconsistency index; and a feedback value analysis submodule, which tracks the correction confirmation information generated by the expert feedback quantification submodule, counts the proportion of it that is adopted by the knowledge increment extraction verification module and finally updated to the graph, and obtains the expert feedback adoption rate.

[0043] Specifically, the coverage calculation submodule sets the statistical period to 30 days. It obtains 60 novel fault cases within this period from the multi-source data dynamic acquisition module via API interface, and then queries the diagnostic records of the closed-loop feedback and inference module, counting 51 successfully diagnosed cases, resulting in a coverage rate of 51 / 60 = 85%. The consistency verification submodule uses stratified sampling, randomly selecting 200 paths from each of the three knowledge dimensions of equipment, materials, and processes, for a total of 600 paths. It applies predefined logical rules such as the inability to have two contradictory standard values ​​for the same parameter. If six contradictory paths are detected, such as "standard value of mold temperature ≤ 60℃" and "standard value of mold temperature ≥ 70℃", the knowledge inconsistency index is 6 / 600 = 1%. The feedback value analysis submodule tracks the entire process of expert feedback through logs. For example, in a certain period, the expert feedback quantification submodule generates 30 correction confirmation messages, of which 24 are verified by the knowledge increment extraction verification module and updated to the graph, resulting in an expert feedback adoption rate of 24 / 30 = 80%. After summarizing the calculation results of each submodule, a health report is generated, including trend curves and improvement suggestions. For example, when the inconsistency index exceeds 2%, it is recommended to perform graph consistency cleaning. This implementation method enables refined evaluation of knowledge graph quality and improves the targeted nature of system maintenance.

[0044] In summary, this invention solves the problem of static knowledge graphs being unable to incorporate new changes by using a multi-source data dynamic acquisition module to capture new fault cases and expert experience in real time. The knowledge increment extraction verification and adaptive update module dynamically expands the ontology, incrementally updates the graph, and ensures data consistency, guaranteeing that the graph can incorporate fault knowledge related to new materials and equipment in real time, thus expanding the fault coverage. The closed-loop feedback and reasoning module combines expert feedback to optimize diagnostic results and improve diagnostic accuracy. The knowledge graph health assessment module quantitatively monitors graph quality, providing direction for system optimization. In conclusion, this system achieves dynamic iteration of the knowledge graph, effectively improving the timeliness and accuracy of fault diagnosis and meeting the dynamic needs of actual production.

[0045] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0046] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A knowledge graph-based injection molding process fault diagnosis system, characterized in that, The system includes: The multi-source dynamic data acquisition module is used to collect injection molding process data, which includes new fault case data, maintenance and solution records, expert experience and feedback, and new equipment specifications. The module performs data preprocessing and structuring, and outputs a multi-source dynamic data stream. The knowledge increment extraction and verification module, based on the multi-source dynamic data stream, performs entity recognition and relation extraction, performs rule learning and pattern discovery, detects conflicts between newly added knowledge and existing knowledge graphs, completes preliminary knowledge verification, and obtains incremental knowledge units to be integrated. The knowledge graph adaptive update module receives the incremental knowledge units to be integrated, performs ontology mapping and alignment operations, calls the incremental graph update algorithm, adds, deletes, and modifies nodes in the injection molding process knowledge graph, records graph version information, and generates an updated injection molding process knowledge graph. The closed-loop feedback and reasoning module, based on the updated injection molding process knowledge graph and combined with real-time fault symptoms, performs reasoning to generate diagnostic results. It receives expert correction and confirmation information on the diagnostic results through a human-computer interaction interface, and transmits the correction and confirmation information as expert experience and feedback to the multi-source data dynamic acquisition module.

2. The injection molding process fault diagnosis system based on knowledge graphs according to claim 1, characterized in that, The multi-source dynamic data stream includes structured fault logs, unstructured maintenance report texts, expert feedback rule sets, and standardized process parameter tables; the incremental knowledge units to be integrated include new fault entities, new causal triples, new diagnostic rules, and extended ontology concepts; the updated injection molding process knowledge graph includes updated equipment component nodes, process parameter edges, fault phenomenon attributes, and solution paths; the correction confirmation information includes diagnostic result correctness labels, knowledge graph node correction suggestions, and newly added experience rule texts.

3. The injection molding process fault diagnosis system based on knowledge graphs according to claim 1, characterized in that, The multi-source data dynamic acquisition module includes: The real-time data capture submodule is used to access the equipment operation log system and alarm record database, capture new fault case data in real time, filter fault records not covered by the knowledge graph, and generate fault log sequences. The semi-structured information parsing submodule calls the fault log sequence to parse the operation steps, replaced parts and adjusted parameters in the maintenance and solution records, extracts key information fields, and generates structured maintenance records; The expert knowledge acquisition submodule receives expert experience and feedback through a human-computer interaction interface, converts the correction opinions or new rules described in natural language into a predefined format, and generates expert knowledge data. The standardized document processing submodule is used to process technical manuals and process specification documents for new equipment and materials, extract key performance parameters and operational constraints, and generate standardized equipment specification data.

4. The injection molding process fault diagnosis system based on knowledge graphs according to claim 1, characterized in that, The knowledge increment extraction and verification module includes: The hybrid entity relation extraction submodule, based on the multi-source dynamic data stream, uses natural language processing technology to perform named entity recognition on unstructured text, identifies entities such as process parameters, fault phenomena, and solutions, and uses a relation extraction model to extract the associations between entities and generate candidate knowledge triples. The association rule mining submodule analyzes sensor readings and fault records in structured data, uses machine learning algorithms to discover potential causal relationships and fault modes between data, and generates candidate diagnostic rules. The knowledge conflict detection submodule compares the candidate knowledge triples and candidate diagnostic rules with the injection molding process knowledge graph to detect logical conflicts, factual contradictions, and data redundancy, and marks conflicting knowledge points. The knowledge rationality verification submodule calls the preset physical constraint rules and domain knowledge model to conduct a preliminary rationality assessment of candidate knowledge that does not conflict, selects the knowledge that passes the verification, and obtains the incremental knowledge units to be integrated.

5. The injection molding process fault diagnosis system based on knowledge graphs according to claim 4, characterized in that, The hybrid entity relationship extraction submodule uses a pre-trained language model based on a transformer to process the maintenance report text, identify the fault cause entity and solution entity contained therein, and calculate the semantic association strength between the two. When the semantic association strength exceeds a preset association threshold, a fault-cause-solution relationship triplet is generated.

6. The injection molding process fault diagnosis system based on knowledge graphs according to claim 1, characterized in that, The knowledge graph adaptive update module includes: The ontology dynamic extension submodule analyzes whether the new entity or new relationship exceeds the definition range of the current ontology based on the incremental knowledge unit to be integrated. If it does, the ontology structure is dynamically extended, new concept classes or attribute definitions are added, and the extended ontology is generated. The incremental update execution submodule calls the extended ontology to map the entities and relations in the incremental knowledge units to be integrated to graph nodes and edges. It uses incremental graph database transaction operations to add nodes, modify attributes, and delete edges to ensure the atomicity of the update process. The version control and rollback submodule records a snapshot of the knowledge graph's state and a change log before and after each update operation. It assigns a unique version number to each successful update and supports rolling back the knowledge graph to a specified historical state based on the version number.

7. The injection molding process fault diagnosis system based on knowledge graphs according to claim 6, characterized in that, The incremental update execution submodule adopts a write-ahead logging mechanism. Before actually modifying the map data file, the update operation is recorded in the log file to ensure that data consistency can be restored by replaying the log in the event of a system failure, thus ensuring the reliability of the update.

8. The injection molding process fault diagnosis system based on knowledge graphs according to claim 1, characterized in that, The closed-loop feedback and inference module includes: The multi-hop inference diagnosis submodule receives real-time fault symptoms as input, performs multi-hop graph traversal in the updated injection molding process knowledge graph, searches for the shortest or highest confidence path from the fault phenomenon node to the root cause node, and generates a preliminary diagnosis result containing the inference path. The diagnostic results presentation submodule visualizes the preliminary diagnostic results on the human-computer interaction interface, clearly presenting the fault phenomenon, the possible causes, and the corresponding solution suggestions. The expert feedback quantification submodule provides experts with scoring of diagnostic results, correction text boxes, and interfaces for adding new knowledge entries. It transforms unstructured feedback input by experts into structured correction and confirmation information, which includes instructions for adjusting the confidence level of specific triples in the graph.

9. The injection molding process fault diagnosis system based on knowledge graphs according to claim 1, characterized in that, The system also includes a knowledge graph health assessment module; The knowledge graph health assessment module periodically evaluates the quality of the injection molding process knowledge graph, calculates knowledge coverage, knowledge redundancy, knowledge consistency and expert feedback adoption rate, and generates a health assessment report. The knowledge coverage rate represents the proportion of fault cases that the current knowledge graph can explain to the total fault case library. The expert feedback adoption rate represents the proportion of expert feedback that has been integrated into incremental knowledge units to the total number of expert feedback.

10. The injection molding process fault diagnosis system based on knowledge graphs according to claim 9, characterized in that, The knowledge graph health assessment module includes: The coverage calculation submodule counts the number of new fault case data collected by the multi-source data dynamic acquisition module within a certain period of time that can be successfully diagnosed by the closed-loop feedback and reasoning module, calculates the ratio of this number to the total number of new cases, and obtains the knowledge coverage rate. The consistency verification submodule randomly extracts knowledge paths from the knowledge graph, applies predefined logical rules to perform reasoning, detects whether there are contradictory reasoning results, calculates the proportion of contradictory paths to the total sampled paths, and obtains the knowledge inconsistency index. The feedback value analysis submodule tracks the correction confirmation information generated by the expert feedback quantification submodule, calculates the proportion of the information that is adopted by the knowledge increment extraction and verification module and finally updated in the graph, and obtains the expert feedback adoption rate.

Images