A method and system for multi-channel similarity identification and duplicate determination of quality event instances

By generating multi-granularity event profiles and a hierarchical candidate recall mechanism, combined with multi-channel similarity calculation, the problem of difficulty in identifying duplicate events and unclear judgment results caused by the various recording forms of quality events in the GxP industry has been solved. This has enabled rapid and accurate identification and judgment of similar events, meeting GxP compliance requirements.

CN123115808APending Publication Date: 2026-11-17CHONGQING AKSO INFORMATION TECH CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611129737.2
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-07-28
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing technologies in the GxP industry struggle to effectively identify and handle various recording formats of quality events, non-standard descriptions, difficulties in identifying duplicate events due to large data volumes, incomplete recall of similar events, lack of interpretability of judgment results, and high computational complexity.

Method used

By generating multi-granular event profiles with multi-level features, combining a hierarchical candidate recall mechanism and multi-channel similarity calculation, the candidate event set is compressed, and interpretable duplicate judgment results are provided through multi-channel similarity calculation and evidence package display.

Benefits of technology

It enables rapid identification and determination of duplicate or similar events in large-scale QMS data, reduces computational costs, improves identification accuracy, provides interpretable evidence packages, and meets the compliance requirements of GxP environments.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123115808A_ABST
    Figure CN123115808A_ABST
Patent Text Reader

Abstract

This invention discloses a method and system for multi-channel similarity identification and duplicate determination of quality event instances. The method includes: receiving quality event instances and determining whether they meet triggering conditions; performing field standardization processing and extracting quality business entities and key phrases; generating multi-granularity event profiles and lightweight fingerprints; employing a hierarchical recall mechanism to recall candidate events from a historical event database; performing multi-channel similarity calculations on the current event and candidate events one by one, generating similarity scores for each channel and calculating a weighted comprehensive similarity; classifying similarity levels based on the comprehensive similarity and generating duplicate determination suggestions and evidence packages, which are displayed on the user interface; receiving user processing actions and writing them into the manual feedback object and audit trail, incrementally updating the event profile and index. This invention solves the problems of difficult identification, slow recall, and lack of interpretability of duplicate or similar events under large-scale QMS data while ensuring recognition accuracy.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of GxP compliance information systems and data similarity retrieval technology, and in particular to a method and system for multi-channel similarity identification and duplicate determination of quality event instances. Background Technology

[0002] In GxP industries such as pharmaceuticals, medical devices, and life sciences, Quality Management Systems (QMS) are typically used to manage records of deviations, complaints, changes, CAPA (Capsule for Actions on Pollution), audit findings, laboratory anomalies, and supplier quality incidents. However, in real-world business operations, a quality incident may enter the system in multiple forms. For example, it may be reported separately by different departments, triggered by records from different batches or equipment, simultaneously triggered by customer complaints and internal deviations, or formed into multiple instances by importing offline records, emails, attachments, and laboratory data.

[0003] For similarity identification of such data, existing technologies mainly fall into four categories: First, precise matching based on key fields, which is simple to implement but cannot handle situations such as missing fields, abbreviations, aliases, typos, or date offsets; Second, inverted indexes and statistical methods such as BM25 based on keywords or full-text retrieval, which can cover some text overlap scenarios, but mainly rely on literal matching and are insufficient in identifying semantically similar but drastically different descriptions such as "bottling machine jam" and "equipment bottle conveying abnormality causing shutdown"; Third, similarity algorithms that introduce edit distance, cosine similarity, or general text vectors, which improve the limitations of pure keyword matching, generally lack constraints on QMS business semantics such as event type, factory, batch, process step, root cause classification, and life cycle status, making it easy to misjudge two types of "temperature exceeding limit" events occurring in the cold chain and sterilization process as similar, while missing actual problems with different descriptions under the same batch and the same equipment; Fourth, relying on manual retrieval and experience judgment by QA personnel, which not only has unstable processing standards but is also difficult to audit and trace.

[0004] Meanwhile, quality event data is characterized by continuous growth. If a textual or semantic comparison of the entire historical record is performed every time a new or updated event is created, the time complexity and space overhead will be high as the data volume continues to grow, affecting the system response speed when creating, saving, or initiating investigation processes. Furthermore, existing technologies generally lack interpretable evidence at the field or fragment level, and manual judgment results are often only used as process notes and are not transformed into reusable rule parameters or training samples. This means that subsequent similarity identification still needs to be calculated from scratch, and it is impossible to continuously correct and optimize the identification effect using historical judgments.

[0005] Therefore, how to ensure that QMS can understand the semantic tasks of quality events, quickly narrow down the candidate range, and provide interpretable duplicate determination results has become a technical problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] To address the aforementioned shortcomings of existing technologies, this invention provides a method and system for multi-channel similarity identification and duplicate determination of quality event instances. By generating multi-level feature multi-granularity event profiles, combining a hierarchical candidate recall mechanism to compress the candidate event set, and performing multi-channel similarity calculation with integrated business constraints, this invention solves the problems of difficulty in identifying duplicate or similar events, slow recall, and lack of interpretability under large-scale QMS data while ensuring identification accuracy.

[0007] To solve the above-mentioned technical problems, the present invention adopts the following technical solution:

[0008] A method for multi-channel similarity identification and duplicate determination of quality event instances includes the following steps:

[0009] S1. Receive quality event instances from QMS and determine whether the current operation meets the similarity recognition trigger condition;

[0010] S2. When it is determined that the similarity recognition triggering condition is met, the structured fields and text fields of the quality event instance are standardized, and quality business entities and key phrases are extracted from the standardized fields to generate standardized event data and entity set.

[0011] S3. Based on the standardized event data and entity set, generate a multi-granularity event profile of the quality event instance, and generate multiple lightweight fingerprints based on the multi-granularity event profile;

[0012] S4. Based on the multi-granularity event profile and lightweight fingerprint, a hierarchical recall mechanism is used to recall candidate events from the historical event database, and the recall results of each level are merged and deduplicated to form a candidate event set.

[0013] S5. Perform multi-channel similarity calculation on the current quality event instance and each candidate event in the candidate event set one by one to generate a similarity score for each channel, and calculate the comprehensive similarity based on the similarity score of each channel and its corresponding weight.

[0014] S6. Based on the comprehensive similarity, the candidate events are divided into multiple similarity levels, and duplicate judgment suggestions are generated. Based on the similarity scores of each channel, an evidence package is generated for the candidate events, and the duplicate judgment suggestions, similarity level division results and evidence packages are displayed on the QMS front-end interface.

[0015] S7. Receive the user's processing actions for the duplicate judgment suggestions and similarity level classification results, write the processing actions into the manual feedback object and audit trail, and incrementally update the event profile and index according to the status changes of the quality event instance.

[0016] As a preferred option, in step S1, the similarity recognition triggering conditions include creating and saving, submitting for approval, changes in key fields, entering the investigation stage, and import completion.

[0017] As a preferred embodiment, in step S2, the field standardization process includes: standardizing the format of product name, batch number, equipment number, material code, supplier name and factory name; standardizing date, production shift and process stage; and explicitly marking null values, unknown values ​​and conflicting values.

[0018] As a preferred embodiment, in step S3, the multi-granularity event profile includes record-level profile, field-level profile, fragment-level profile, causal chain profile, and relation-level profile; the lightweight fingerprint includes structured field hash, key entity SimHash, text MinHash, phrase fingerprint, causal chain fingerprint, and relational neighborhood fingerprint.

[0019] As a preferred embodiment, in step S4, the processing procedure of the hierarchical recall mechanism includes:

[0020] First, partition recall and blocking key recall are performed to narrow down the historical event database to the first intermediate candidate size. Then, fingerprint recall and inverted index recall are performed to narrow down the first intermediate candidate size to the second intermediate candidate size. Finally, approximate vector recall and relation graph recall are performed to narrow down the second intermediate candidate size to the candidate event set.

[0021] The size of the candidate event set is smaller than the size of the second intermediate candidate, the size of the second intermediate candidate is smaller than the size of the first intermediate candidate, and the size of the first intermediate candidate is smaller than the historical event database.

[0022] As a preferred embodiment, in step S5, the comprehensive similarity is the weighted sum of the similarity scores of each channel. The weights corresponding to each channel are configured according to the event type using a business constraint weighting mechanism, and the calculation formula is expressed as follows:

[0023] ;

[0024] In the formula: The overall similarity between the current quality event instance A and the candidate event B; For structured field similarity; For literal text similarity; Semantic vector similarity; For quality entity similarity; For causal chain similarity; For relational graph similarity; For temporal and spatial similarity; For lifecycle context similarity; to Configurable weights for each similarity channel.

[0025] As a preferred embodiment, in step S5, the business constraint weighting mechanism is as follows: when any structured field of the quality event instance is missing, the similarity weight of the channel corresponding to the missing field is reduced; when the quality event instance and the candidate event conflict on key fields, a conflict penalty is triggered to reduce the overall similarity.

[0026] As a preferred embodiment, in step S6, the evidence package includes field hits, text fragment hits, entity hits, relation object hits, and similarity channel scores and corresponding weights for each channel.

[0027] When the candidate event is not visible to the current user, the de-identified summary is displayed or the candidate event is not displayed according to the preset permission configuration.

[0028] As a preferred embodiment, in step S7, the incremental update of the event profile and index based on the state changes of the quality event instance specifically includes: incrementally maintaining the event profile, inverted index, fingerprint index, vector index, and relationship graph index when the quality event instance is created, updated, archived, deleted, or its state changes; and performing cold and hot tiered storage for closed or archived events.

[0029] A multi-channel similarity identification and duplicate determination system for quality event instances, used to execute the multi-channel similarity identification and duplicate determination method for quality event instances as described above, including:

[0030] The quality event data access module is used to receive quality event instances in QMS and determine whether the current operation meets the similarity recognition trigger conditions.

[0031] The event standardization and entity normalization module is used to standardize the structured fields and text fields of the quality event instance, and extract quality business entities and key phrases from the standardized fields to generate standardized event data and entity sets.

[0032] The multi-granularity event profile generation module is used to generate multi-granularity event profiles of the quality event instances based on the standardized event data and entity set, and to generate multiple lightweight fingerprints based on the multi-granularity event profiles;

[0033] The hierarchical candidate recall module is used to recall candidate events from the historical event database based on the multi-granularity event profile and lightweight fingerprint, and to merge and deduplicate the recall results of each level to form a candidate event set.

[0034] The multi-channel similarity calculation module is used to perform multi-channel similarity calculation on the current quality event instance and each candidate event in the candidate event set one by one, generate a similarity score for each channel, and calculate the comprehensive similarity based on the similarity score of each channel and its corresponding weight.

[0035] The duplicate determination and display module is used to divide the candidate events into multiple similarity levels based on the comprehensive similarity, generate duplicate determination suggestions, generate evidence packages for the candidate events based on the similarity scores of each channel, and display the duplicate determination suggestions, similarity level division results and evidence packages on the QMS front-end interface.

[0036] The index incremental maintenance module is used to receive user actions on the duplicate judgment suggestions and similarity level classification results, write the actions into the manual feedback object and audit trail, and incrementally update the event profile and index according to the status changes of the quality event instance.

[0037] Compared with the prior art, the present invention has the following technical effects:

[0038] 1. This invention standardizes the structured and text fields of quality event instances and extracts quality business entities and key phrases to generate multi-granularity event profiles. It integrates structured fields, unstructured text, business entities, object relationships, and causal chains into multi-level feature representations. On this basis, multi-channel similarity calculation calculates similarity from eight dimensions: structured fields, literal text, semantic vectors, entities, causal chains, relationship graphs, time-space, and lifecycle context. This enables the system to identify the same or similar events under different input methods.

[0039] 2. This invention designs a multi-level progressive hierarchical recall mechanism, which gradually compresses the historical event database into a candidate event set through low-cost recall methods. This allows multi-channel similarity calculations to be performed only on the candidate event set. In addition, it incrementally updates event profiles and indexes based on the status changes of quality event instances, and performs cold and hot tiered storage for closed or archived events. This significantly reduces the computational cost of a single identification task and maintains a stable response speed even when the data volume continues to grow, avoiding high-cost pairwise comparisons of all historical events.

[0040] 3. This invention decomposes quality event text into phenomenon fragments, impact fragments, cause fragments, handling fragments, and prevention fragments through fragment-level profiling and causal chain profiling. This allows for the calculation of similarity scores for each channel by separately calculating the similarity of each fragment and then determining the overall similarity of the link structure, thereby distinguishing different business scenarios. Simultaneously, when configuring the weights of each channel, the weight is configured according to the event type. When any structured field is missing, the weight of the channel corresponding to the missing field is reduced. When there is a conflict with a candidate event on a key field, a conflict penalty is triggered to reduce the overall similarity. This allows the judgment logic to be flexibly adjusted according to different scenarios such as deviations, complaints, and CAPA, thereby reducing false positives and false negatives from the root.

[0041] 4. This invention generates an evidence package containing field hits, text fragment hits, entity hits, relation object hits, and similarity scores and corresponding weights for each channel. This allows users to clearly understand the contribution of each channel to the overall similarity score, reducing the threshold for manual review. At the same time, by writing the user-confirmed processing actions such as duplication, similarity, and non-duplication into the manual feedback object, the manual judgment conclusions are precipitated as reusable data for subsequent recall ranking, rule weight adjustment, or model optimization. The interpretability of the evidence package and the continuous accumulation of manual feedback objects form a closed loop, enabling the similarity recognition capability to continuously evolve in actual use.

[0042] 5. This invention incorporates lifecycle context similarity into multi-channel similarity calculation, avoiding repeated prompts for confirmed duplicates or excluded events. Furthermore, when a candidate event is not visible to the current user, a de-identified summary is displayed or the candidate event is not displayed based on permission configuration, ensuring that similarity retrieval does not violate data isolation policies. Simultaneously, user actions are written into the audit trail, supporting the tracing of differences in judgments of the same event at different times. The combination of these three aspects enables this invention to be deployed compliantly in a GxP-controlled environment, meeting audit traceability requirements. Attached Figure Description

[0043] To make the objectives, technical solutions, and advantages of the invention clearer, the invention will now be described in further detail with reference to the accompanying drawings, wherein:

[0044] Figure 1 This is a flowchart of the multi-channel similarity identification and duplicate determination method for quality event instances proposed in this embodiment of the invention;

[0045] Figure 2 This is a system architecture diagram of the multi-channel similarity recognition and duplicate determination system for quality event instances proposed in this embodiment of the invention;

[0046] Figure 3 This is a diagram showing the relationship between similarity determination results and evidence objects in an embodiment of the present invention. Detailed Implementation

[0047] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0048] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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 a part of the embodiments of the present invention, not all of them. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to represent selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0049] Example:

[0050] In GxP industries such as pharmaceuticals, medical devices, and life sciences, Quality Management Systems (QMS) are typically used to manage records of deviations, complaints, changes, CAPA (Capsule for Assessment and Approval), audit findings, laboratory anomalies, and supplier quality incidents. A quality incident may enter the system in various forms during real-world business operations. For data similarity identification, existing technologies mainly fall into four categories: 1) deduplication based on precise field matching; 2) similarity queries based on keywords or full-text search; 3) general text similarity algorithms based on edit distance, cosine similarity, or text vectors; and 4) manual retrieval and judgment. However, these existing technologies generally suffer from the following problems: First, field matching and keyword retrieval rely on literal overlap, failing to identify semantically similar but differently worded events such as "bottling machine jam" and "equipment bottle conveying abnormality causing machine shutdown." Second, general text similarity algorithms lack the semantic constraints of QMS business operations. For example, even with the same "temperature exceeding limit," the nature and handling procedures of the event are completely different in cold chain warehousing and sterilization processes, but text similarity might misjudge them as similar; while events with different wording under the same batch and the same equipment may be missed. Third, comparing each new event with every record in the entire historical data would incur a computational cost close to O(N) or O(N²), which would severely impact system response speed as the data volume increases. Fourth, the system only outputs a "similarity of 85%" score, lacking field-level or fragment-level evidence, making it difficult for users to accept and hindering audit trails; the results of manual judgments are also not retained as reusable data, requiring subsequent identification to start from scratch, thus preventing continuous optimization.

[0051] To address the aforementioned problems and shortcomings, this invention proposes a multi-channel similarity identification and duplicate judgment method and system for quality event instances. It aims to solve the problems in existing quality management systems (QMS) such as difficulty in identifying duplicate events, incomplete recall of similar events, and lack of interpretability of judgment results caused by diverse event entry methods, non-standard descriptions, and large data volumes.

[0052] The invention will now be further described in three parts: system composition, data object and field design, and method flow.

[0053] I. Specific Implementation Process of the Method

[0054] The multi-channel similarity identification and duplicate determination method for quality event instances proposed in this invention, such as... Figure 1 As shown, it includes the following steps:

[0055] S1. Receive quality event instances from QMS and determine whether the current operation meets the similarity recognition trigger condition;

[0056] The steps in this embodiment are the starting point of the similarity identification process. The system monitors the lifecycle changes of newly created, updated, or imported quality event instances (such as deviations, complaints, CAPA, audit findings, etc.) in QMS in real time through the data access module.

[0057] Triggering conditions: Similarity identification is not performed every time data changes, but is driven by preset triggering conditions. According to this embodiment, the similarity identification triggering conditions include: creation and saving, submission for approval, change of event title or description, change of product or batch number, change of equipment field, change of root cause field, entry into the investigation stage, and completion of import, etc.

[0058] S2. When it is determined that the similarity recognition triggering condition is met, the structured fields and text fields of the quality event instance are standardized, and quality business entities and key phrases are extracted from the standardized fields to generate standardized event data and entity set.

[0059] In this embodiment, when a quality event is created, saved, submitted, has its key fields modified, or enters the investigation phase, the structured and text fields of the quality event are first standardized using the event standardization and entity normalization module. This standardization process includes:

[0060] Standardize the format: Standardize the format of batch number, equipment number, material code, product name, supplier name, and factory name, and normalize aliases, abbreviations, differences in capitalization, differences in full-width and half-width characters, and mixed Chinese and English characters into standard entities;

[0061] Entity identification: Identify quality business entities from the text of the title, description, root cause, impact assessment, and CAPA description of the quality event instance. The quality business entities include equipment, products, materials, process steps, inspection items, defect phenomena, root cause classification, and CAPA actions.

[0062] Date and Stage Standardization: Standardize the occurrence date, discovery date, production shift, and process stage;

[0063] Null value marking: Explicitly mark null, unknown, and conflicting values ​​instead of discarding them directly.

[0064] Simultaneously, quality business entities and key phrases are identified from the text of the title, description, root cause, impact assessment, and CAPA description of quality event instances. These quality business entities include equipment, products, materials, process steps, inspection items, defect phenomena, root cause classifications, and CAPA actions. The standardized data and entity set serve as the basic input for subsequent processing.

[0065] S3. Based on the standardized event data and entity set, generate a multi-granularity event profile of the quality event instance, and generate multiple lightweight fingerprints based on the multi-granularity event profile;

[0066] In this embodiment, a quality event is comprehensively characterized by constructing a multi-dimensional event profile. This event profile includes the following:

[0067] Record-level profile: Composed of key fields such as event type, product, batch number, equipment, factory, date of occurrence, and severity, used to determine the overall characteristics of the event.

[0068] Field-level profiling: Generate text features, semantic vectors, and entity lists for fields such as title, event description, root cause, impact assessment, and CAPA description.

[0069] Fragment-level profiling: Long texts are broken down into phenomenon fragments, impact fragments, cause fragments, treatment fragments, and prevention fragments in order to identify local similarities.

[0070] Causal chain profile: Organize event information according to the structure of "abnormal phenomenon - affected object - possible cause - confirmed cause - corrective measures - preventive measures".

[0071] Relationship-level profiling: Extracting the relationships between events and objects such as batches, equipment, suppliers, documents, CAPA, changes, and complaints.

[0072] Simultaneously, this module generates multiple lightweight fingerprints for events, including structured field hashes, key entity SimHashes, text MinHashes, phrase fingerprints, causal chain fingerprints, and relational neighborhood fingerprints. Different fingerprints are used in different recall stages, enabling the system to quickly match similar events at different granularities.

[0073] The above technical solution integrates the structured fields, unstructured text, business entities, object relationships, and causal chains of quality events into a multi-level feature representation through a multi-granularity event profiling mechanism, enabling the system to capture similar features of events from different levels; lightweight fingerprints provide efficient matching basis for rapid recall and avoid misjudgment caused by missing fields, abbreviations, aliases, or format differences.

[0074] S4. Based on the multi-granularity event profile and lightweight fingerprint, a hierarchical recall mechanism is used to recall candidate events from the historical event database, and the recall results of each level are merged and deduplicated to form a candidate event set.

[0075] In this embodiment, to address the problem that full pairwise comparisons in the prior art are costly as the data volume increases, this embodiment designs a hierarchical candidate recall mechanism. This mechanism does not directly perform detailed calculations on each current event and all historical events, but instead first narrows down the candidate range quickly through multiple recall channels.

[0076] Specifically, the process of this tiered recall mechanism includes:

[0077] Level 1: Partition Recall and Blocking Key Recall: The system determines the retrieval partition based on conditions such as event type, factory, product line, time range, and lifecycle status, and recalls historical events within the corresponding partition. Simultaneously, it generates blocking keys based on combinations of fields such as product, batch number, equipment, supplier, process step, and defect type, recalling records with the same or similar blocking keys. Through this level, the system narrows the historical event database from N entries to a first intermediate candidate size of N1.

[0078] Level 2: Fingerprint Recall and Inverted Index Recall: The system quickly identifies records with similar literal meanings or similar entity sets using SimHash, MinHash, phrase fingerprints, or field hashes; it recalls text-related historical events using keywords, standard entities, quality terms, and phrase indexes. Through this level, the system narrows down the first intermediate candidate size N1 to the second intermediate candidate size N2.

[0079] Level 3: Approximate Vector Recall and Relationship Graph Recall: The system recalls historical events with similar semantic vectors using an approximate nearest neighbor index (such as the HNSW algorithm); and recalls business-related historical records based on object relationships such as the same batch, same device, same CAPA, same supplier, same audit findings, and same document. Through this level, the system narrows down the second intermediate candidate size N2 to the final candidate event set K.

[0080] Where K is much smaller than N2, N2 is much smaller than N1, and N1 is much smaller than N. In addition, the system also has a historical judgment recall channel, which prioritizes recalling relevant records that have been manually identified as duplicates, similarities, or non-duplicates to assist in the current judgment.

[0081] The above technical solution uses a hierarchical recall mechanism. The system first compresses the candidate range quickly with low cost, and then progressively narrows the candidate set. This allows subsequent fine similarity calculations to be performed only on candidate sets that are much smaller than the entire database. The computational complexity of a single recognition task is reduced from O(N) to O(K). Even when quality events continue to increase, the system can still maintain a stable response speed, avoiding user waiting and resource bottlenecks caused by pairwise comparisons across the entire database after the data volume expands.

[0082] S5. Perform multi-channel similarity calculation on the current quality event instance and each candidate event in the candidate event set one by one to generate a similarity score for each channel, and calculate the comprehensive similarity based on the similarity score of each channel and its corresponding weight.

[0083] In this embodiment, to address the problem in existing technologies that rely solely on single text similarity while ignoring the QMS business context, this embodiment calculates multiple sub-similarity scores between the current quality event instance and each candidate event in the candidate event set using multi-channel similarity calculation, as shown in Table 1, including:

[0084] Table 1. Similarity of each channel

[0085] Similarity Channel Example computed object effect Structured field similarity Event type, product, batch number, equipment, factory, department, date window Determine if the business facts are close Title and description literal similarity Title, Problem Description, Preliminary Investigation Identify similarities in keywords, phrases, and text morphology. Semantic vector similarity Description, root causes, impact assessment, CAPA text vectors Identify events with different literal meanings but similar semantics. Quality Entity Similarity Products, equipment, materials, defects, process steps, root cause classification Determine whether the quality business elements overlap or are similar. Causal chain similarity The link between phenomenon, impact, cause, and measures Determine if the logical structures of events are similar Relationship graph similarity Batch, equipment maintenance, changes, CAPA, complaints, audit findings Determine if an event shares upstream or downstream objects. temporal and spatial similarity Factory, production line, shift, time window Determine whether the facts may have originated from the same scene. Lifecycle context similarity Process status, manual judgment, historical merging relationships Avoid repeatedly prompting for records that have been processed or excluded.

[0086] After the similarity of each channel is calculated, the system calculates the overall similarity using the following formula:

[0087] ;

[0088] In the formula: The overall similarity between the current quality event instance A and the candidate event B; For structured field similarity; For literal text similarity; Semantic vector similarity; For quality entity similarity; For causal chain similarity; For relational graph similarity; For temporal and spatial similarity; For lifecycle context similarity; to Configurable weights are assigned to each similarity channel. In this embodiment, the weights can be adjusted based on event type, quality process, field completeness, business risk level, and customer configuration. For example, for deviation events, the channel corresponding to product, batch number, equipment, process step, abnormal phenomenon, and occurrence time has a higher weight than other channels; for complaint events, the channel corresponding to product, batch number, complaint phenomenon, customer, and defect type has a higher weight than other channels; for CAPA events, the channel corresponding to root cause classification, action type, responsible department, recurrence pattern, and related events has a higher weight than other channels.

[0089] Meanwhile, the system incorporates a business constraint weighting mechanism: when any structured field of a quality event instance is missing, the system automatically reduces the similarity weight of the channel corresponding to the missing field; when a quality event instance and a candidate event conflict on key fields, the system triggers a conflict penalty, reducing the overall similarity. This prevents a single channel from abnormally dominating the final result and also avoids misjudging duplication based solely on text similarity.

[0090] S6. Based on the comprehensive similarity, the candidate events are divided into multiple similarity levels, and duplicate judgment suggestions are generated. Based on the similarity scores of each channel, an evidence package is generated for the candidate events, and the duplicate judgment suggestions, similarity level division results and evidence packages are displayed on the QMS front-end interface.

[0091] In this embodiment, after the comprehensive similarity calculation is completed, the candidate events are divided into the following five levels based on the comprehensive similarity and key channel results:

[0092] Suspected duplication: The overall similarity exceeds the first threshold, and the key business fields or causal chain meet the duplication judgment conditions.

[0093] High similarity: The overall similarity is high, but there is a lack of sufficient evidence to directly determine it as a duplicate.

[0094] Historical recurrence reference: Events may not be repeated, but they are similar in root cause, equipment, process steps, defect type, or CAPA failure mode.

[0095] Low similarity references: Local similarity information exists, which users can expand to view.

[0096] No duplicates or excluded: It should not be indicated as a duplicate after confirmation by rules or manual verification.

[0097] Based on these five levels, duplicate determination suggestions are generated for each candidate event, and the results are provided to the user interface, workflow, notification, or subsequent data association.

[0098] The above technical solution employs multi-channel similarity calculation across eight dimensions, enabling the system to simultaneously identify literal and semantic similarity. A causal chain similarity mechanism distinguishes between "similar phenomena," "similar root causes," "similar measures," and "repetition of the same event." A business constraint weighting mechanism allows the judgment logic to be flexibly adjusted according to different scenarios such as deviations, complaints, and CAPA (Capsule for Actions on Problems). The similarity hierarchy classification transforms the judgment result from a vague single score into clear, actionable business recommendations. These combined measures fundamentally reduce false positives and false negatives.

[0099] S7. Receive the user's processing actions for the duplicate judgment suggestions and similarity level classification results, write the processing actions into the manual feedback object and audit trail, and incrementally update the event profile and index according to the status changes of the quality event instance.

[0100] In this embodiment, to address the problem of insufficient interpretable evidence in similarity results in the prior art, this embodiment generates an evidence package for each candidate event, including:

[0101] The hit structured fields include, for example, same product, same batch number, same equipment, same factory, and occurrence date difference of 2 days.

[0102] The matched text fragments, such as those containing keywords like "filling," "bottleneck," and "shutdown" in both event descriptions, or those semantic fragments that both point to "production interruption caused by abnormal delivery."

[0103] The quality entities that are hit include, for example, the same equipment, the same process step, and the same defect type.

[0104] The matched related objects, such as two events that are both associated with the same CAPA or the same equipment maintenance order.

[0105] The scores and weights of each similarity channel.

[0106] The system determines the level and trigger threshold.

[0107] User-generated confirmation, exclusion, merging, or association of operation records.

[0108] At the same time, the system synchronously outputs the contribution decomposition of the comprehensive similarity, allowing users to clearly see how much the structured fields, semantic text, and relation graphs contributed to the comprehensive score, and to further drill down to view the specific fields and textual evidence corresponding to each contribution item.

[0109] The above technical solution changes the way the system in GxP scenarios only outputs a single score of "85% similarity". Users and auditors usually need to know which fields, text fragments, business entities, and historical relationships the similarity comes from, and why the system judged it as "suspected duplication" or "similar but not duplicate". Without field-level or fragment-level evidence, users are unlikely to accept the results, which is also not conducive to subsequent audit traceability. This embodiment enables QA personnel to quickly understand and accept the system's judgment, reduces the threshold and cost of manual review, and facilitates audit traceability.

[0110] In this embodiment, to address the problem that duplicate determination results are not retained as reusable data in the prior art, this embodiment allows users to process system prompts through manual confirmation and feedback learning. Processing actions include confirming duplicates, confirming similar but not duplicates, establishing associations, ignoring, excluding, merging events, and creating historical recurrence descriptions.

[0111] User actions generate similarity judgment records, which are written to a manual feedback object. Key fields in this object include the source event ID, candidate event ID, user conclusion, processing reason, electronic signature information, audit record reference, and whether it enters the rule optimization sample. For the GxP system, manual confirmation results should retain the operator, time, reason, results before and after the change, electronic signature, or audit trail information. If the customer environment does not allow for automatic model learning, the system can also use manual judgments only as rule parameters and audit data, without directly updating the algorithm model.

[0112] The aforementioned technical solution transforms user-confirmed conclusions regarding duplication, similarity, and non-duplication into reusable data assets for subsequent recall ranking, rule weight adjustment, or model optimization. The interpretability of evidence and the continuous accumulation of human feedback form a closed loop, enabling the system's similarity recognition capabilities to iteratively evolve in practical use.

[0113] In this embodiment, to address the problems of high cost of full index reconstruction and heavy pressure on historical data storage in the prior art, this embodiment designs an incremental index maintenance module. When a quality event is created, updated, archived, deleted, or its status changes, the event profile, inverted index, fingerprint index, vector index, and relationship graph index are incrementally maintained.

[0114] For closed or archived events, the system uses a hot / cold tier storage approach based on time, object type, event type, or tenant: active events retain complete profiles, vectors, and evidence; recently closed events retain commonly used indexes and compressed vectors; and long-archived events retain summary profiles, quantization vectors, and necessary relational indexes. Through incremental indexing, the system avoids recalculating the features of all historical events for each query and avoids pairwise comparisons across the entire database. This allows for faster recall even as the number of high-quality events continues to grow, while reducing the storage pressure from long-term historical data.

[0115] In this embodiment, to address the compliance requirements for data isolation and audit traceability in GxP scenarios, this embodiment uses audit trails and access control to record information such as the trigger time, input events, candidate recall channels, similarity calculation version, threshold version, result display, and user actions for similarity identification tasks. Simultaneously, this module controls whether users can view certain historical events or similar evidence based on user roles, object permissions, lifecycle status, and data isolation rules. When candidate events involve different factories, departments, or permission domains, the system can display only anonymized summaries or "Restricted similar records exist," preventing users from bypassing business permissions through similarity searches.

[0116] The above technical solution addresses the compliance requirement that "similar searches must not exceed business permissions" through access control; and the audit trail addresses the traceability requirement of "how to explain to auditors why the system prompts or does not prompt for a certain historical event." Together, these two aspects ensure the compliant deployment and reliable operation of this invention in a GxP-controlled environment.

[0117] Furthermore, this embodiment also incorporates comprehensive anomaly and boundary handling mechanisms. When a field is missing, the system explicitly marks the missing field and reduces the corresponding channel weight, rather than treating the missing field as dissimilar or identical. When permissions are restricted, the system displays a de-identification prompt or hides the candidate based on the permission configuration. When the semantic model is unavailable, the system can revert to field, fingerprint, inverted index, and rule similarity without blocking the saving of quality events. Users can mark results as non-duplicate, reducing similar false alarms in the future. The system identifies the main event and the merged event to avoid duplicate prompts. During batch import, the system can asynchronously generate similarity recognition tasks to avoid long waiting times on the front end. In a multi-tenant environment, indexes, profiles, and similarity results are isolated by tenant or customer environment to avoid cross-tenant recall. When the algorithm version changes, the system records the rule version and algorithm version used in each calculation, facilitating the tracing of differences in judgments for the same event at different times. This ensures that the system can still operate stably under various anomalies and boundary conditions, without disrupting normal business processes, while providing complete technical support for auditing and tracing.

[0118] II. Systems for Implementing the Above Methods

[0119] This invention also provides a multi-channel similarity identification and duplicate determination system for quality event instances, comprising multiple cooperating modules for executing the above-described method. For example... Figure 2 As shown, the system includes the following modules:

[0120] The quality event data access module is used to receive quality event instances in QMS and determine whether the current operation meets the similarity recognition trigger conditions.

[0121] The event standardization and entity normalization module is used to standardize the structured fields and text fields of the quality event instance, and extract quality business entities and key phrases from the standardized fields to generate standardized event data and entity sets.

[0122] The multi-granularity event profile generation module is used to generate multi-granularity event profiles of the quality event instances based on the standardized event data and entity set, and to generate multiple lightweight fingerprints based on the multi-granularity event profiles;

[0123] The hierarchical candidate recall module is used to recall candidate events from the historical event database based on the multi-granularity event profile and lightweight fingerprint, and to merge and deduplicate the recall results of each level to form a candidate event set.

[0124] The multi-channel similarity calculation module is used to perform multi-channel similarity calculation on the current quality event instance and each candidate event in the candidate event set one by one, generate a similarity score for each channel, and calculate the comprehensive similarity based on the similarity score of each channel and its corresponding weight.

[0125] The duplicate determination and display module is used to divide the candidate events into multiple similarity levels based on the comprehensive similarity, generate duplicate determination suggestions, generate evidence packages for the candidate events based on the similarity scores of each channel, and display the duplicate determination suggestions, similarity level division results and evidence packages on the QMS front-end interface.

[0126] The index incremental maintenance module is used to receive user actions on the duplicate judgment suggestions and similarity level classification results, write the actions into the manual feedback object and audit trail, and incrementally update the event profile and index according to the status changes of the quality event instance.

[0127] III. System Data Object Design

[0128] To support the operation of the above methods and systems, this embodiment defines the following core data objects in the object model of the QMS or Gaia platform for storing and managing various types of data in the similarity recognition process.

[0129] 1. Objects of Quality Incidents

[0130] This is the main object being compared, namely the deviation, complaint, CAPA, and other records managed in the QMS. Its key fields include: event number, event type, event title, event description, occurrence / discovery date, plant, department, product, material, batch number, equipment, production line, severity, root cause classification, CAPA association, and lifecycle status.

[0131] 2. Event profiling target

[0132] This object stores standardized features extracted and generated from "quality event objects." It forms the basis for fast retrieval and similarity calculation. Its key fields include:

[0133] Event ID: Associated with the original quality event.

[0134] Portrait Version: Used to track the update history of portraits.

[0135] Structured field summary: A standardized set of key fields.

[0136] Standard entity set: A list of business entities extracted from text.

[0137] Text vector reference: Points to the storage address of the semantic vector describing the event.

[0138] Causal chain structure: a structured chain of information consisting of "phenomenon-cause-measure".

[0139] Fingerprint value: Various lightweight fingerprints (such as SimHash, MinHash, etc.) generated for quick retrieval.

[0140] 3. Similarity index objects

[0141] This object is the underlying data structure supporting the hierarchical recall mechanism, used to achieve fast retrieval of massive amounts of data. It is not a single index, but a collection of multiple indexes, including:

[0142] Field blocking key index: constructed based on a combination of fields such as product and batch number.

[0143] Entity / Phrase Inverted Index: Built based on standardized entities and keywords.

[0144] Fingerprint index: such as SimHash bucket, MinHash bucket.

[0145] Semantic vector indexes: such as HNSW, IVF and other approximate nearest neighbor indexes.

[0146] Graph adjacency index: built based on the relationship between events and batches, devices, CAPA and other objects.

[0147] 4. Similarity determination result objects

[0148] This object is used to store the calculation results of each similarity recognition task. Its key fields include: source event ID, candidate event ID, overall similarity, similarity level (e.g., suspected duplication, high similarity), duplication determination suggestion, similarity score of each channel, user processing status, etc.

[0149] 5. Similar evidence objects

[0150] This object is crucial for ensuring the interpretability of the judgment results, and it stores field-level and fragment-level evidence. Its key fields include: result ID, evidence type (field hit, text fragment hit, etc.), source field / text fragment, candidate field / text fragment, hit entity, contribution score, etc.

[0151] 6. Subjects of manual feedback

[0152] This object is used to record the user's actions in response to the system's judgment results, forming the basis for closed-loop optimization. Its key fields include: source event ID, candidate event ID, user conclusion (confirmation of duplication, ignore, etc.), processing reason, electronic signature information, and whether it is used for model optimization.

[0153] like Figure 3 As shown, quality events are associated with similarity tasks, with one event trigger generating one task. Each similarity task contains multiple similarity results, each corresponding to a judgment conclusion for a candidate event. Each similarity result further includes multiple pieces of similarity evidence, recording the judgment basis from four dimensions: fields, text fragments, entities, and relationships. Simultaneously, similarity results are linked to manual feedback through signature verification, recording the user's actions regarding the judgment conclusion. Audit logs are associated with each of these objects, recording the entire process of creation, modification, and operation. This object relationship design ensures complete traceability from identification and triggering to user processing, providing a data foundation for audit trails in GxP compliance scenarios.

[0154] IV. Examples

[0155] To further illustrate the specific applications of the present invention in different business scenarios, four specific embodiments are provided below.

[0156] 1. Example 1: Identification of Duplicate Entry of Deviation Events

[0157] A pharmaceutical company created a deviation event in its QMS (Quality Management System) titled "Bottle jamming at the filling line caused a brief shutdown." The event description recorded that a batch of a certain product experienced bottle jamming during the filling process, resulting in an 8-minute production line downtime. When the user saved the event, the system triggered a similarity check.

[0158] The system first standardizes the events, identifying entities such as product, batch number, filling line, equipment number, process step "filling," and abnormal phenomenon "bottle jam / shutdown." The system then generates structured field fingerprints, title text fingerprints, abnormal phenomenon fragment vectors, and relational neighborhood profiles.

[0159] During the candidate recall phase, the system recalled 12 historical records using the same product, batch number, and equipment blocking key; 30 records using inverted indexes such as "filling," "bottle jam," and "shutdown"; 20 records with different literal meanings but similar semantics, such as "bottle conveying abnormality caused production stoppage," using semantic vectors; and maintenance and deviation records of the same equipment over the past 30 days using equipment relationship graphs. After merging and deduplicating, the system obtained 25 candidate events.

[0160] During the similarity calculation phase, the system found that one candidate event had highly identical product, batch number, equipment, occurrence date, and anomaly, except that the title was "Abnormal conveying of filling equipment." The system gave it a comprehensive similarity score of 91 and marked it as "suspected duplicate." The evidence presentation listed reasons such as the same batch number, the same equipment, the occurrence time difference of 1 hour, and the semantic consistency of the anomaly. After the user confirmed that the event was a duplicate entry, the system recorded the manual confirmation result and prompted the user to choose to keep the main event or establish a connection.

[0161] 2. Example 2: Recommendation of similar quality events between complaints and deviations

[0162] The customer complaint record describes "the batch number on the product outer box as difficult to identify." This complaint may not necessarily be a duplicate record of an internal deviation, but it could be related to the quality of the inkjet printing on the packaging line. The system performs similarity identification when saving the complaint.

[0163] The system extracts entities from complaint texts, including product, batch number, defect "batch number difficult to identify," packaging materials, and customer market. Using semantic vectors and quality entity indexing, the system recalls an internal deviation record titled "Insufficient clarity of outer packaging coding." While the two records are not entirely identical in textual description, they are highly correlated in terms of product, batch number, defect, and packaging process.

[0164] The system flagged it as "highly similar" rather than "suspected duplication" and prompted the user to correlate the complaint with the internal bias. This result can be used in subsequent investigations to cite the root causes of the internal bias and CAPA progress.

[0165] 3. Example 3: CAPA Relapse Risk Identification

[0166] In a deviation survey, the user filled in "device sensor calibration misalignment, causing abnormal temperature control" as the root cause and drafted a CAPA (Conditional Action Plan). The system triggers similarity identification when entering the CAPA drafting phase.

[0167] The system not only retrieves event titles and descriptions, but also generates semantic vectors for root cause fragments and action fragments, and compares them with the root cause classification, action type, and shutdown effect of historical CAPAs. The system found that the same device family had three similar root causes within the past year, and that a similar deviation occurred after one of the CAPAs was shut down.

[0168] The system marks relevant events as "historical recurrence references" and explains in the evidence that similarities mainly stem from root cause classification, equipment family, temperature control links, and measure types. Users can cite these historical events in the investigation report to assess the adequacy of the existing CAPA.

[0169] 4. Example 4: Asynchronous Similarity Recognition After Batch Import of Quality Events

[0170] The company migrated a batch of historical deviation records from the old system to QMS. Due to the large number of records, the system does not perform synchronous calculations for each record in the import interface, but instead creates asynchronous similarity comparison tasks for the import batch.

[0171] The system first generates event profiles and indexes in batches, then performs candidate recall based on event type and time partitioning. For high-risk events or events from the past year, the system calculates complete multi-channel similarity; for long-archived events, the system uses compressed profiles and quantized vectors for recall, performing precise calculations only on high-scoring candidates. After the task is completed, the system generates a list of suspected duplicates and a list of similar events for data administrators to review.

[0172] V. Overview

[0173] In summary, this embodiment addresses the practical problems in QMS quality event management, such as the inability of existing technologies to effectively identify semantic similarity, ignore business context, high cost of full comparison, lack of interpretability of judgment results, and inability to reuse manual judgments. It proposes a method and system for multi-channel similarity identification and duplicate judgment of quality event instances. This embodiment first transforms scattered and inconsistently formatted quality event data into a unified standard representation through an event standardization and entity normalization module, generating multi-granularity event profiles and various lightweight fingerprints to lay the data foundation for subsequent accurate matching. To address the computational pressure brought by large-scale historical data, a hierarchical candidate recall mechanism is designed. Through multi-level progressive compression, the historical event database is reduced from N to K candidate sets. Combined with incremental index maintenance and hot / cold hierarchical storage, real-time responsiveness is maintained even as data continues to grow. In the fine-grained similarity calculation stage, similarity is calculated and weighted across eight dimensions: structured fields, literal text, semantic vectors, quality entities, causal chains, relationship graphs, temporal space, and lifecycle context. In particular, the causal chain similarity mechanism decomposes event text into five segments: phenomenon, impact, cause, measures, and prevention, and judges the degree of proximity of the links. Combined with business constraints such as weighting based on event type, deweighting for missing fields, and penalties for key field conflicts, the judgment logic can be flexibly adjusted according to different scenarios such as deviations, complaints, and CAPA, reducing false positives and false negatives. In addition, the evidence generation and interpretable display module outputs multi-level evidence and contribution decomposition, the manual confirmation and feedback learning module precipitates user conclusions into reusable data to drive continuous optimization, and the audit trail and access control module ensures compliance throughout the entire process.

[0174] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit the technical solutions. Those skilled in the art should understand that any modifications or equivalent substitutions to the technical solutions of the present invention without departing from the spirit and scope of the present invention should be covered within the scope of the claims of the present invention.

Claims

1. A method for multi-channel similarity identification and duplicate determination of quality event instances, characterized in that, Includes the following steps: S1. Receive quality event instances from QMS and determine whether the current operation meets the similarity recognition trigger condition; S2. When it is determined that the similarity recognition triggering condition is met, the structured fields and text fields of the quality event instance are standardized, and quality business entities and key phrases are extracted from the standardized fields to generate standardized event data and entity set. S3. Based on the standardized event data and entity set, generate a multi-granularity event profile of the quality event instance, and generate multiple lightweight fingerprints based on the multi-granularity event profile; S4. Based on the multi-granularity event profile and lightweight fingerprint, a hierarchical recall mechanism is used to recall candidate events from the historical event database, and the recall results of each level are merged and deduplicated to form a candidate event set. S5. Perform multi-channel similarity calculation on the current quality event instance and each candidate event in the candidate event set one by one to generate a similarity score for each channel, and calculate the comprehensive similarity based on the similarity score of each channel and its corresponding weight. S6. Based on the comprehensive similarity, the candidate events are divided into multiple similarity levels, and duplicate judgment suggestions are generated. Based on the similarity scores of each channel, an evidence package is generated for the candidate events, and the duplicate judgment suggestions, similarity level division results and evidence packages are displayed on the QMS front-end interface. S7. Receive the user's processing actions for the duplicate judgment suggestions and similarity level classification results, write the processing actions into the manual feedback object and audit trail, and incrementally update the event profile and index according to the status changes of the quality event instance.

2. The method for multi-channel similarity identification and duplicate determination of quality event instances according to claim 1, characterized in that, In step S1, the similarity recognition triggering conditions include creating and saving, submitting for approval, changes in key fields, entering the investigation stage, and import completion.

3. The method for multi-channel similarity identification and duplicate determination of quality event instances according to claim 1, characterized in that, In step S2, the field standardization process includes: standardizing the format of product name, batch number, equipment number, material code, supplier name and factory name; standardizing date, production shift and process stage; and explicitly marking null values, unknown values ​​and conflicting values.

4. The method for multi-channel similarity identification and duplicate determination of quality event instances according to claim 1, characterized in that, In step S3, the multi-granularity event profile includes record-level profile, field-level profile, fragment-level profile, causal chain profile, and relation-level profile; the lightweight fingerprint includes structured field hash, key entity SimHash, text MinHash, phrase fingerprint, causal chain fingerprint, and relational neighborhood fingerprint.

5. The method for multi-channel similarity identification and duplicate determination of quality event instances according to claim 1, characterized in that, In step S4, the processing procedure of the hierarchical recall mechanism includes: First, partition recall and blocking key recall are performed to narrow down the historical event database to the first intermediate candidate size. Then, fingerprint recall and inverted index recall are performed to narrow down the first intermediate candidate size to the second intermediate candidate size. Finally, approximate vector recall and relation graph recall are performed to narrow down the second intermediate candidate size to the candidate event set. The size of the candidate event set is smaller than the size of the second intermediate candidate, the size of the second intermediate candidate is smaller than the size of the first intermediate candidate, and the size of the first intermediate candidate is smaller than the historical event database.

6. The method for multi-channel similarity identification and duplicate determination of quality event instances according to claim 1, characterized in that, In step S5, the overall similarity is a weighted sum of the similarity scores of each channel. The weights corresponding to each channel are configured according to the event type using a business constraint weighting mechanism, and the calculation formula is as follows: ; In the formula: The overall similarity between the current quality event instance A and the candidate event B; For structured field similarity; For literal text similarity; Semantic vector similarity; For quality entity similarity; For causal chain similarity; For relational graph similarity; For temporal and spatial similarity; For lifecycle context similarity; to Configurable weights for each similarity channel.

7. The method for multi-channel similarity identification and duplicate determination of quality event instances according to claim 6, characterized in that, In step S5, the business constraint weighting mechanism is as follows: when any structured field of the quality event instance is missing, the similarity weight of the channel corresponding to the missing field is reduced. When the quality event instance and the candidate event conflict on a key field, a conflict penalty is triggered to reduce the overall similarity.

8. The method for multi-channel similarity identification and duplicate determination of quality event instances according to claim 1, characterized in that, In step S6, the evidence package includes field matching, text fragment matching, entity matching, relation object matching, and similarity channel scores and corresponding weights for each channel; When the candidate event is not visible to the current user, the de-identified summary is displayed or the candidate event is not displayed according to the preset permission configuration.

9. The method for multi-channel similarity identification and duplicate determination of quality event instances according to claim 1, characterized in that, In step S7, the incremental update of the event profile and index based on the state changes of the quality event instance specifically includes: incrementally maintaining the event profile, inverted index, fingerprint index, vector index, and relationship graph index when the quality event instance is created, updated, archived, deleted, or its state changes; and performing cold and hot tiered storage for closed or archived events.

10. A multi-channel similarity recognition and duplicate determination system for quality event instances, characterized in that, A method for performing multi-channel similarity identification and duplicate determination of quality event instances as described in any one of claims 1 to 9, comprising: The quality event data access module is used to receive quality event instances in QMS and determine whether the current operation meets the similarity recognition trigger conditions. The event standardization and entity normalization module is used to standardize the structured fields and text fields of the quality event instance, and extract quality business entities and key phrases from the standardized fields to generate standardized event data and entity sets. The multi-granularity event profile generation module is used to generate multi-granularity event profiles of the quality event instances based on the standardized event data and entity set, and to generate multiple lightweight fingerprints based on the multi-granularity event profiles; The hierarchical candidate recall module is used to recall candidate events from the historical event database based on the multi-granularity event profile and lightweight fingerprint, and to merge and deduplicate the recall results of each level to form a candidate event set. The multi-channel similarity calculation module is used to perform multi-channel similarity calculation on the current quality event instance and each candidate event in the candidate event set one by one, generate a similarity score for each channel, and calculate the comprehensive similarity based on the similarity score of each channel and its corresponding weight. The duplicate determination and display module is used to divide the candidate events into multiple similarity levels based on the comprehensive similarity, generate duplicate determination suggestions, generate evidence packages for the candidate events based on the similarity scores of each channel, and display the duplicate determination suggestions, similarity level division results and evidence packages on the QMS front-end interface. The index incremental maintenance module is used to receive user actions on the duplicate judgment suggestions and similarity level classification results, write the actions into the manual feedback object and audit trail, and incrementally update the event profile and index according to the status changes of the quality event instance.