Aircraft product quality problem management zeroing method and system based on structured evidence chain

Abstract

This invention provides a method and system for zeroing out quality issues in aviation products based on a structured evidence chain, belonging to the field of aviation product quality engineering and reliability management technology. The method includes the following steps: S1, constructing a structured rule knowledge base: breaking down the five stages of management zeroing into multiple secondary sub-stages, defining the necessary evidence items and types for each secondary sub-stage, forming a mapping table of stages, evidence items, and types, and obtaining a structured rule knowledge base; S2, preprocessing the evidence and performing triple automated verification; S3, constructing a visualized evidence chain: after verification, generating a dynamic and interactive visualized evidence chain graph. This invention solves the problem of incomplete zeroing caused by incomplete evidence chains and non-closed logical relationships in existing management zeroing processes, improving review efficiency by more than 98%, and reducing the probability of recurrence of similar problems to below 5%. It is applicable to zeroing out scenarios for quality issues in aviation products such as military aircraft and civil passenger aircraft.

Description

Technical Field

[0001] This invention relates to the field of aviation product quality engineering and reliability management technology, specifically to a method and system for zeroing out aviation product quality problems based on a structured chain of evidence. Background Technology

[0002] In the development of high-tech equipment such as aerospace, zeroing out quality issues is a crucial step in preventing recurrence, improving the effectiveness of the management system, and ensuring the reliability of equipment development. Relevant national military standards, such as GJB / Z194-2021, have clearly defined the general requirements and implementation process for zeroing out quality issues. The core of this process includes five key steps: "identifying the problem's occurrence, identifying responsibility, taking measures, handling the issue seriously, and improving regulations," referred to as the five steps of zeroing out quality issues. These steps provide principled guidance for the work of zeroing out quality issues.

[0003] However, in actual engineering applications and implementation, the existing management zeroing model still has many prominent technical problems, making it difficult to effectively implement standard requirements and ensuring the depth and quality of management zeroing. Specifically, this is reflected in the following three aspects: (1) The execution process is overly formalistic and lacks substantive depth: It only clarifies the goal and principle requirements of management zeroing, but fails to clearly define the specific forms of evidence, content specifications, and logical connections that must be output in each of the five stages of management zeroing. When compiling zeroing reports, each research and development unit relies heavily on the personal experience of its staff in terms of evidence selection, content detail, and logical argumentation. This leads to a large number of zeroing reports falling into the trap of piling up text and listing regulations, failing to truly construct a complete logical closed loop from problem root cause investigation and responsibility determination to long-term improvement measures. The management zeroing process becomes a mere formality and difficult to implement. For example, in the zeroing work on the problem of blade cracks in a certain type of aero-engine, the relevant units only submitted the blade crack detection report, without linking it to key supporting materials such as operation records of the corresponding process and personnel qualification certificates. This made it impossible to trace the management-level causes of the cracks, and the zeroing work failed to achieve substantive results.

[0004] (2) Fragmented and disordered evidence management, making logical verification difficult: The entire zeroing process generates a large amount of various types of evidence, including audit records, liability determination letters, penalty documents, training records, rectification implementation documents, etc. Currently, such evidence is mostly stored separately as independent files on different carriers such as document servers and local hard drives, lacking a unified associated index and standardized management. During manual review, reviewers need to search for various scattered evidence across systems and folders, making it difficult to systematically verify the supporting relationships, citation relationships, causal relationships, and corroborative relationships between different pieces of evidence. It is very easy to miss key logical breakpoints, such as liability determination conclusions not referencing corresponding job responsibility documents, rectification measures not being associated with the root cause analysis results, etc., directly leading to the inability to effectively guarantee the rigor and scientific nature of the zeroing conclusion.

[0005] (3) Low review efficiency and difficulty in reusing successful experiences: When reviewing the zeroing-out report, the quality supervisor needs to manually check the completeness of the evidence at each stage, such as whether the measures taken are complete, including the rectification plan, implementation records, effect verification results, standardization, such as whether the document signatures are complete, whether the format meets the requirements, and the logic, such as whether the rectification measures are accurately aimed at the root cause of the problem, and whether the responsibility determination matches the problem occurrence process. The average review time for a single case is 4-6 hours, which is not only a large workload and time-consuming, but also prone to oversights due to subjective factors. In addition, the effective experience and standardized evidence chains contained in past successful management zeroing-out cases are all saved in unstructured document formats such as Word and PDF, which cannot be quickly retrieved, accurately matched and learned from based on problem characteristics such as welding defects, equipment inspection loopholes, and improper material selection. This leads to low implementation efficiency of drawing inferences from one instance to another, and the probability of the same quality problem recurring is as high as 30% or more, which violates the core original intention of management zeroing-out.

[0006] In summary, the existing technology lacks a technical means to structurally decompose the standard requirements of management zeroing, automatically verify the compliance of evidence throughout the zeroing process, and forcibly construct and systematically verify the logical relationships between various types of evidence. As a result, the depth, standardization, and effectiveness of the management zeroing process cannot be effectively guaranteed, making it difficult to meet the stringent quality control requirements of the aerospace high-tech equipment development field. Therefore, a technical solution that can solve the above-mentioned technical problems is urgently needed. Summary of the Invention

[0007] To address the shortcomings of existing technologies, this invention aims to provide a zero-out method and system for aviation product quality issue management based on a structured evidence chain. This addresses the problems of formalized processes, low efficiency, and knowledge silos caused by non-standardized evidence, incomplete logical loops, and reliance on manual review in existing zero-out processes. It provides a systematic technical solution that enforces the rigor of the zero-out process, achieves efficient automated review, and promotes knowledge accumulation and reuse. The core is the construction of a rule-driven, automatically verified, and logically visible quality issue management zero-out auxiliary system. The system includes a user access layer, a core business application layer, a data support layer, and a lower-level technical support layer. The hardware architecture includes client terminals, application servers, database servers, and AI computing nodes. The software architecture adopts a layered design, including a presentation layer, a business logic layer, a data layer, and an AI engine layer.

[0008] Specifically, on the one hand, the present invention provides a method for zeroing out quality problems in aviation products based on a structured chain of evidence, which includes the following steps: S1. The five-step management zeroing process is broken down into multiple second-level sub-steps. The necessary evidence items and their data types are defined for each second-level sub-step. Then, the ternary attribute requirements are defined for each evidence item, and the logical relationship rules between different evidence items are defined and stored in the rule knowledge base. S2. Preprocess and perform triple automated verification on the evidence submitted in each sub-step of the management zeroing process, specifically: S21. Preprocess the evidence and convert multimodal evidence into a structured format; S22. Perform triple automated verification on the evidence. The triple automated verification includes normative verification, integrity verification, and logical relationship verification. The logical relationship verification is specifically as follows: S221. Evidence Entity Extraction: A trained BERT-BiLSTM-CRF model is used to extract core entities from each piece of evidence text. The BERT-BiLSTM-CRF model includes a BERT encoding layer, a BiLSTM temporal feature extraction layer, and a CRF label decoding layer. The BERT-BiLSTM-CRF model employs an improved sequence label loss function, which, based on the standard conditional random field loss, integrates entity boundary penalty terms and label transition constraint terms to enhance its ability to model the boundaries of zero-valued evidence entities and label dependencies. The improved sequence label loss function is as follows: ; Where S() is the loss function, and N is the number of training samples. As a sample, Let i be the true label sequence of the i-th sample. For all possible label sequences; S222. Construct an entity association graph: Construct an entity association graph based on entity similarity and contextual association; S223. Verify Logical Relationships: A matching algorithm based on graph isomorphism and rule embedding is used to match the entity association graph with the logical relationship rules in the rule knowledge base. Specifically, each logical relationship rule is converted into a graph pattern, where nodes are entity types and edges are relationship types. For the entity association graph G and the rule pattern P, the VF2 subgraph isomorphism algorithm is used to calculate whether there is a subgraph in G that is isomorphic to P. If so, the match is successful; otherwise, the maximum common subgraph is calculated, and the matching degree = |common subgraph| / |P| is output. A first matching degree threshold and a second matching degree threshold are preset and increased sequentially. When the matching degree is greater than or equal to the second matching degree threshold, the logical relationship is considered valid. When the matching degree is greater than or equal to the first matching degree threshold and less than the second matching degree threshold, the logical relationship is considered partially matched. When the matching degree is less than the first matching degree threshold, a logical breakpoint is identified. S3. Construct a visual evidence chain: After verification, a dynamic and interactive visual evidence chain graph is generated.

[0009] Preferably, step S1 specifically includes the following sub-steps: S11. The five-step management zeroing process is broken down into 23 secondary sub-steps according to the management zeroing standard process, and each secondary sub-step corresponds to at least one management action. S12. Define the necessary evidence items and their data types for each secondary sub-stage, forming a stage-evidence item-type mapping table. Data types include document type, image type, structured data type, and audio type. S13. Define a ternary attribute requirement for each piece of evidence. The ternary attribute includes normative requirements, completeness requirements, and validity requirements. S14. First-order predicate logic is used to define the logical relationship rules between different evidence items. The logical relationship rules include supporting relationships, referencing relationships, causal relationships and corroborating relationships, and are stored in the rule knowledge base.

[0010] Preferably, in step S14, based on the characteristics of the zero-chain evidence, each logical relationship is further subdivided into several subcategories, totaling 12 core logical relationships. Specifically, supporting relationships are divided into direct support, indirect support, and evidence chain support; citation relationships are divided into document citation, clause citation, and external standard citation; causal relationships are divided into direct causation, indirect causation, and conditional causation; and corroboration relationships are divided into witness testimony-documentary evidence corroboration, physical evidence-record corroboration, and time corroboration. Furthermore, 58 subdivided rules are obtained and stored in the relational rule base of the rule engine. The relational rule base is a component of the structured rule knowledge base.

[0011] Preferably, the preprocessing of evidence in step S21 specifically involves: extracting text content and metadata for document-type evidence using Apache Tika; performing clarity detection and text recognition for image-type evidence using OpenCV; verifying field integrity for structured data-type evidence using the Python Pandas library; and converting audio-type evidence to a standard format using FFmpeg, while simultaneously generating a text transcript.

[0012] Preferably, in step S22, the normativity verification is based on the normativity attribute requirements in the rule knowledge base, and the normativity is verified in two dimensions using metadata verification and content verification, and the verification result is output; the integrity verification is based on the integrity attribute requirements in the rule knowledge base, and the integrity of evidence items and attachments is verified, and the verification result is output.

[0013] Preferably, in step S221, the BERT encoding layer uses a pre-trained BERT model to encode the input evidence text sequence, obtaining the context-related vector representation of each character. The input text sequence is... ,in Let t be the t-th character, and T be the sequence length. After passing through the BERT encoder, the output is the hidden layer vector of each character. , dimension , For the BERT hidden layer dimension, the specific expression is: ; in, These are the trainable parameters of the BERT model.

[0014] Preferably, in step S221, the BiLSTM temporal feature extraction layer extracts the context vector output by BERT. The input is a Bidirectional Long Short-Term Memory (BiLSTM) network to capture the forward and backward temporal dependencies of a text sequence. The BiLSTM network consists of a forward LSTM and a backward LSTM. The forward LSTM processes the sequence from left to right and outputs the forward hidden layer vector. The inverse LSTM processes the sequence from right to left and outputs the inverse hidden layer vector. The two are concatenated to obtain the output vector of BiLSTM. , dimension Forward LSTM computation: ; ; ; ; The reverse LSTM computation is symmetrical to the forward one, obtained by concatenating the forward and reverse hidden layer vectors. ,in The activation vectors for the input gate, forget gate, and output gate, respectively. Let be the cell state vector. This is the weight matrix. For bias vectors, For activation function, It is an element-wise product.

[0015] Preferably, in step S221, the CRF tag decoding layer introduces a Conditional Random Field (CRF) for tag decoding, and the CRF layer outputs a BiLSTM signal. As input, it is first mapped to a label score matrix through a fully connected layer. , dimension , This represents the score for the t-th character labeled as the k-th class label: ; That This is the weight matrix of the fully connected CRF layer. As the bias vector, the goal of the CRF layer is to find the optimal label order. To maximize the sequence score, the sequence score is calculated using the following formula: ; in, This represents the transition weight from label i to label j. The transition matrix, As the starting tag, To terminate the labeling process, reasonable transition relationships between labels are learned through training. The Viterbi algorithm is used to solve for the optimal label sequence. : .

[0016] Preferably, the specific steps of step S222 are as follows: Entity similarity calculation: Cosine similarity is used to calculate the semantic similarity between text entities. Combined with attribute similarity, candidate entity pairs with similarity higher than the similarity threshold θ_sim are selected. Contextual association mining: Constructing association edges based on the temporal, spatial, and logical attributes of evidence; Graph construction: Using each extracted entity as a node and similarity or association as an edge, an initial entity association graph G(V,E) is generated. Isolated nodes and weak association edges with weights lower than the weight threshold θ_edge are removed to obtain the final entity association graph.

[0017] On the other hand, the present invention provides a zeroing system for a zeroing method for managing quality problems in aviation products based on a structured evidence chain, which includes a structured rule knowledge base construction module, an automated verification module, and a visual evidence chain construction module. The structured rule knowledge base construction module is used to break down the five-stage management zeroing process into multiple secondary sub-stages, define the necessary evidence items and types for each secondary sub-stage, form a mapping table of stages, evidence items and types, clearly define the three-element attribute requirements for each evidence item and define the logical relationship rules between different evidence items, thus obtaining the structured rule knowledge base; The automated verification module is used for preprocessing evidence and performing triple automated verification. The visual evidence chain building module is used to generate a dynamic and interactive visual graph of the evidence chain after verification.

[0018] Compared with the prior art, the beneficial effects of the present invention are as follows: The zeroing-out system for aviation product quality issue management based on a structured evidence chain provides an approach to resolving issues. By constructing a structured rule knowledge base, it transforms the abstract requirements of the "five stages of management zeroing-out" into computable rules containing multimodal evidence items, ternary attributes, and logical relationships. Relying on a guided process and a triple automated verification engine (standardization, completeness, and logical relationships), it achieves unified parsing and compliance verification of multimodal evidence. The logical relationship verification, based on the BERT-BiLSTM-CRF model and first-order predicate logic, automatically detects breakpoints in causal and referential relationships between evidence. Simultaneously, it constructs a dynamic and visualized evidence chain, generates a structured zeroing-out report, and enables intelligent reuse of cases through multi-dimensional tagged storage. This invention solves the problems of formalization, low review efficiency, and knowledge silos in existing management zeroing-out processes, improving review efficiency by over 98% and reducing the probability of recurring similar issues to below 5%. It is applicable to zeroing-out scenarios for aviation product quality issue management in military aircraft, civil passenger aircraft, and other aviation products.

[0019] The present invention provides a zeroing-out method for aviation product quality problem management based on a structured evidence chain. Through a structured rule base, the zeroing-out work that relies on personal experience is transformed into a standardized operation process, ensuring the consistency of the execution process for different units and different personnel.

[0020] (3) The zeroing-out method for aviation product quality problem management based on structured evidence chain provided by this invention can discover logical breakpoints that are easily overlooked by manual review by automatically verifying the logical relationship between evidence. It can forcefully ensure a complete closed loop from problem identification to measure implementation and system improvement, and significantly improve the depth and rigor of the zeroing-out report. Automated verification greatly reduces the workload of quality supervisors, and the review efficiency can be improved by more than 98%, allowing them to focus on the technical decision itself. At the same time, the systematic inspection avoids human oversight and greatly improves the review efficiency and quality.

[0021] (4) In the zeroing method for aviation product quality problem management based on structured evidence chain provided by the present invention, each zeroing case and its evidence chain are stored in a structured manner, which facilitates intelligent retrieval based on semantics or problem features, provides strong data support for “learning from one case and applying it to others” and preventing similar problems, transforms personal experience into organizational assets, and can realize the structured accumulation and reuse of knowledge, and has a wide range of application scenarios in engineering practice. Attached Figure Description

[0022] Figure 1 This is a schematic diagram of the overall process of the present invention; Figure 2 This is a block diagram of the overall architecture of the zeroing system for the zeroing method for managing aviation product quality problems based on structured evidence chains, as described in this invention. Figure 3 This is a flowchart of the zeroing-out method for aviation product quality problem management based on structured evidence chain, as described in this invention. Figure 4 This is a schematic diagram illustrating the structure of the structured rule knowledge base of the present invention; Figure 5 This is a visual representation of the evidence chain in an exemplary serious handling process according to an embodiment of the present invention. Figure 6 This is a flowchart of the automated logical relationship verification of the present invention. Detailed Implementation

[0023] Hereinafter, embodiments of the present invention will be described with reference to the accompanying drawings.

[0024] Specifically, on the one hand, this invention provides a method for zeroing out quality problems in aviation products based on a structured chain of evidence, such as... Figure 1 , Figure 2 and Figure 3 As shown, it includes the following steps: S1. Construct a structured rule knowledge base: The five-step management zeroing process is broken down into multiple secondary sub-steps. For each secondary sub-step, necessary evidence items and their data types are defined. Then, ternary attribute requirements are defined for each evidence item, and logical relationship rules between different evidence items are defined and stored in the rule knowledge base. This step specifically includes the following sub-steps: S11. Deconstructing the Zeroing-Out Process: The five-step management zeroing-out process is broken down into 23 secondary sub-steps. For example, "identifying the process of the problem" is broken down into four sub-steps: "process audit evidence collection", "execution record collection", "problem phenomenon physical evidence collection", and "relevant personnel interview record collection". Each secondary sub-step corresponds to at least one management action.

[0025] S12. Define the necessary evidence items and their data types for each secondary sub-stage, forming a stage-evidence item-type mapping table. Data types include document types, image types, structured data types, and audio types. Among them, evidence types include document types such as PDF / Word, image types such as JPG / PNG with a resolution ≥300dpi, structured data types such as Excel / database tables with field integrity ≥95%, and audio types such as MP3 / WAV with a sampling rate ≥44.1kHz, including text transcripts.

[0026] S13. Define ternary attribute requirements for each piece of evidence. These ternary attributes include normative requirements, completeness requirements, and validity requirements. Normative requirements include document format, signature requirements, and field format; completeness requirements include the number of required fields and the completeness of attachments; and validity requirements include time validity and source validity. Normative requirements include document format (e.g., PDF / A-3a), signature requirements (e.g., electronic signatures must comply with GB / T 35273-2020), and field format (e.g., date format is YYYY-MM-DD); completeness requirements include the number of required fields (e.g., a liability determination letter must include 8 required fields such as the responsible person, type of liability, and basis for determination), and the completeness of attachments (e.g., training records must include attendance sheets, courseware, and grade sheets); and validity requirements include time validity (e.g., the evidence generation time must be after the problem occurred but before the zeroing process is completed) and source validity (e.g., audit records must be issued by the company's audit department).

[0027] S14. Formalization of Logical Relationship Rules: First-order predicate logic is used to define the logical relationship rules between different evidence items. These rules include supporting relationships, referencing relationships, causal relationships, and corroborating relationships, totaling 12 core logical relationships and 58 sub-rules, stored in the relationship rule base of the rule engine. The logical relationship rules include four basic relationships: supporting relationships, referencing relationships, causal relationships, and corroborating relationships. Based on this, according to the characteristics of zero-chain evidence, each basic relationship is further subdivided into several subcategories, resulting in a total of 12 core logical relationships. For example: supporting relationships are subdivided into "direct supporting," "indirect supporting," and "chain of evidence supporting"; referencing relationships are subdivided into "document referencing," "clause referencing," and "external standard referencing"; causal relationships are subdivided into "direct causality," "indirect causality," and "conditional causality"; and corroborating relationships are subdivided into "witness testimony-documentary evidence corroboration," "physical evidence-record corroboration," and "time corroboration," etc., deriving 58 sub-rules, which are stored in the relationship rule base of the rule engine. The relationship rule base is a component of the structured rule knowledge base in step S1 and is specifically used to store the logical relationship rules between evidence items.

[0028] In a specific embodiment, 12 core logical relationships and 58 detailed rules are defined, stored in the relational rule base of the rule engine. These relationships include: support relationships (e.g., "Responsibility Determination Letter (B3)" supports "Disciplinary Notice (D1)", with the predicate expression Support(B3,D1)); reference relationships (e.g., "Disciplinary Notice (D1)" references "Responsibility Determination Decision (B3)", with the predicate expression Quote(D1,B3)); causal relationships (e.g., "Equipment Inspection Missing Information (A1)" leads to "Undetected Cold Welding (A3)", with the predicate expression Cause(A1,A3)); and corroboration relationships (e.g., "Interview Record (A4)" corroborates "Inspection Log (A2)", with the predicate expression Confirm(A4,A2)). Table 1 below shows an example of the zeroing-evidence item-type mapping table. Table 2 shows an example of the ternary attribute requirements for evidence items.

[0029] Table 1: Zeroing-out Stage - Evidence Item - Type Mapping Table (Example)

[0030] Table 2: Tripartite Attribute Requirements for Evidence Items (Example: Liability Determination Decision B3)

[0031] S2. Preprocess and perform triple automated verification on the evidence submitted in each sub-step of the zeroing process, specifically including the following sub-steps: S21. Preprocess the evidence, converting multimodal evidence into a uniform, processable structured format. Specifically, the preprocessing involves: extracting text content and metadata from document-type evidence using Apache Tika; performing sharpness detection and character recognition on image-type evidence using OpenCV; verifying field integrity of structured data-type evidence using the Python Pandas library; and converting audio-type evidence to a standard format using FFmpeg, while simultaneously generating a text transcript.

[0032] In practical applications, a wizard-style interface is used. Each secondary sub-step displays three functional areas: "List of Evidence Items," "Attribute Requirements Hints," and "Sample File Download." For example, the "Process Audit Evidence Collection" sub-step interface displays "Required Submission: Special Audit Report" on the left, "Normative Requirements: PDF format, including auditor's signature" in the middle, and provides a sample audit report that meets the requirements for download on the right. After users upload evidence, the system automatically performs preprocessing: document evidence is extracted using Apache Tika to extract text content and metadata (such as author, creation time, and signature information); image evidence is processed using OpenCV for clarity detection (variance ≥100 is considered acceptable) and text recognition (OCR, using the PaddleOCR engine, with a recognition accuracy ≥98%); structured data evidence is verified for field integrity using the Python Pandas library (missing fields are automatically highlighted in red); audio evidence is converted to a standard format using FFmpeg and simultaneously uses the Baidu speech recognition API to generate a text transcript. The system displays the progress of evidence submission at each stage in real time (e.g., "Investigation of the problem: 4 / 4 pieces of evidence have been submitted, 100%)". Stages that do not meet the standards are marked in red, and the system also indicates the name of the evidence item that has not been submitted and the reason (e.g., "Interview transcript not submitted: Please upload a scanned copy with the interviewee's signature").

[0033] S22. Perform triple automated verification on the evidence. The triple automated verification is normative verification, integrity verification and logical relationship verification.

[0034] The normative verification is based on the normative attribute requirements in the rule knowledge base. It uses metadata verification and content verification to perform a two-dimensional verification of normativity and outputs the verification results. In a specific embodiment, Java POI / Adobe PDFLibrary is used to read document metadata and verify file format, signature information, and field format. Keyword matching is performed on the document text content (e.g., the responsibility determination decision must contain keywords such as "responsibility type" and "basis for determination"; if missing, the message "content is incomplete, responsibility type description needs to be supplemented") is displayed. Field extraction is performed on the image OCR results (e.g., photos of faulty parts must contain text information such as "serial number" and "inspection date"; if not extracted, the message "image needs to be labeled with product serial number" is displayed). Verification result output: A "Normative Verification Report" is generated, marking qualified / unqualified items. Unqualified items must specify the reasons and rectification suggestions (e.g., "Unqualified item: Audit report has no signature; Reason: No electronic signature conforming to GB / T 35273-2020 was detected; Rectification suggestion: Supplement the quality department's electronic signature and re-upload").

[0035] The integrity verification is based on the integrity attribute requirements in the rule knowledge base. It uses the integrity of evidence items and the integrity of attachments to perform integrity verification and outputs the verification results. In a specific embodiment, by comparing the evidence items submitted by the user with the "list of required evidence items" for this stage, it is determined whether there are any missing items (e.g., the "ascertaining responsibility" stage requires 4 pieces of evidence, and if only 3 pieces are submitted, it will prompt "Missing 'Responsibility Determination Notification Receipt (B4)', please supplement it").

[0036] The "deduction system for missing items" is adopted. 25 points will be deducted for each missing required evidence item and 10 points will be deducted for each missing attachment. The maximum score is 100 points. 80 points or above is considered passing. If the score is below 80 points, the submission must be revised and resubmitted.

[0037] like Figure 6 As shown, the logical relationship verification specifically includes the following sub-steps: S221. Evidence Entity Extraction: A trained BERT-BiLSTM-CRF model is used to extract core entities from each piece of evidence text. The BERT-BiLSTM-CRF model includes a BERT encoding layer, a BiLSTM temporal feature extraction layer, and a CRF label decoding layer. Addressing the characteristics of ambiguous entity boundaries and strong label dependencies in zero-level evidence for quality issue management (e.g., the entity "responsible person" must appear in a specific paragraph of the "responsibility determination letter"), the BERT-BiLSTM-CRF model of this invention employs an improved sequence label loss function. This function, based on the standard Conditional Random Field (CRF) loss, integrates an entity boundary penalty term and a label transition constraint term to enhance the modeling ability of entity boundaries and label dependencies in zero-level evidence. The final loss function is as follows: ; Where S() is the loss function, and N is the number of training samples. As a sample, Let i be the true label sequence of the i-th sample. Let S represent all possible label sequences. The loss function S() is the cross-entropy loss.

[0038] The BERT encoding layer uses a pre-trained BERT model to encode the input evidence text sequence, obtaining a context-related vector representation for each character. The input text sequence is... ,in Let t be the t-th character, and T be the sequence length. After passing through the BERT encoder, the output is the hidden layer vector of each character. , dimension , For the BERT hidden layer dimension, the specific expression is: ; in, These are the trainable parameters of the BERT model.

[0039] The BiLSTM temporal feature extraction layer extracts the context vector output by BERT. The input is a Bidirectional Long Short-Term Memory (BiLSTM) network to capture the forward and backward temporal dependencies of a text sequence. The BiLSTM network consists of a forward LSTM and a backward LSTM. The forward LSTM processes the sequence from left to right and outputs the forward hidden layer vector. The inverse LSTM processes the sequence from right to left and outputs the inverse hidden layer vector. The two are concatenated to obtain the output vector of BiLSTM. , dimension Forward LSTM computation: ; ; ; ; The reverse LSTM computation is symmetrical to the forward one, obtained by concatenating the forward and reverse hidden layer vectors. ,in The activation vectors for the input gate, forget gate, and output gate, respectively. Let be the cell state vector. This is the weight matrix. For bias vectors, For activation function, It is an element-wise product.

[0040] Preferably, the CRF tag decoding layer introduces a Conditional Random Field (CRF) for tag decoding, and the CRF layer outputs a BiLSTM signal. As input, it is first mapped to a label score matrix through a fully connected layer. , dimension , This represents the score for the t-th character labeled as the k-th class label: ; That This is the weight matrix of the fully connected CRF layer. As the bias vector, the goal of the CRF layer is to find the optimal label order. To maximize the sequence score, the sequence score is calculated using the following formula: ; in, This represents the transition weight from label i to label j. The transition matrix, As the starting tag, To terminate the labeling process, reasonable transition relationships between labels are learned through training. The Viterbi algorithm is used to solve for the optimal label sequence. : .

[0041] The BERT-BiLSTM-CRF model used in this invention does not directly use existing general models. Instead, it makes the following improvements to address the domain-specific characteristics of zero-quality evidence texts (such as up to 15 entity types and strong domain constraints on label transfer): During the BERT pre-training stage, more than 100,000 pieces of aviation quality zero-quality domain corpus are injected for domain-adaptive fine-tuning; in the CRF layer, a domain constraint matrix is ​​introduced to prohibit illegal label transfers (such as the "responsible person" label cannot be directly followed by the "time" label); and an entity boundary penalty term is added to the loss function.

[0042] S222. Constructing Entity Relationships: Based on entity similarity and contextual relationships, construct an entity relationship graph. The specific steps are as follows: Entity similarity calculation: Cosine similarity (based on BERT embedding vectors) is used to calculate the semantic similarity between text entities. At the same time, attribute similarity is combined (e.g., the edit distance between the job titles “AOI Operator” and “Optical Inspector” is ≤2) to filter out candidate entity pairs with similarity higher than the threshold θ_sim (default 0.75).

[0043] Contextual association mining: Construct association edges based on the time attributes of evidence (e.g., the generation time difference between the evidence of two entities is ≤7 days), spatial attributes (e.g., the same process, the same equipment number), and logical attributes (e.g., one entity appears in the "basis" field of another entity).

[0044] Graph construction: Using each extracted entity as a node and similarity or association as an edge (edge ​​weight = α·similarity + β·context matching degree, α + β = 1), generate an undirected / directed entity association graph G(V,E).

[0045] Graph pruning: Remove isolated nodes (degree=0) and weakly related edges with weights below the threshold θ_edge (default 0.3) to obtain the final entity association graph.

[0046] S223. A matching algorithm based on graph isomorphism and rule embedding is used to match the entity association graph with the logical relationship rules in the rule knowledge base, specifically including: Formalizing the rules: Each logical relationship rule (such as Cause(entity A, entity B)) is converted into a graph pattern, where nodes are entity types and edges are relationship types.

[0047] Matching degree calculation: For an entity association graph G and a rule pattern P, the VF2 subgraph isomorphism algorithm (optimized with node label and edge type constraints) is used to calculate whether there exists a subgraph in G that is isomorphic to P. If it exists, the match is successful; otherwise, the size of the largest common subgraph is calculated, and the matching degree = |common subgraph| / |P| is output.

[0048] When the matching degree is greater than or equal to the second matching degree threshold, the logical relationship is determined to be valid; when the matching degree is greater than or equal to the first matching degree threshold and less than the second matching degree threshold, the logical relationship is determined to be partially matched; when the matching degree is less than the first matching degree threshold, it is determined to be a logical breakpoint.

[0049] In a specific embodiment, the first matching threshold is 0.5, and the second matching threshold is 0.85. When the matching degree is ≥0.85, the logical relationship is determined to be valid; when the matching degree is ≤0.5 and <0.85, it is marked as "partial match" and a prompt is given indicating that it needs to be improved; when it is <0.5, it is marked as "logical breakpoint".

[0050] Causal verification: For example, extract the problem entity "Equipment Inspection Omission" from "Audit Report (A1)" and the abnormal entity "Calibration Status Field Blank" from "Detection Log (A2)". Verify the causal relationship using the causal relationship rule (Cause(problem entity, abnormal entity)). If the two descriptions are consistent and the time intervals overlap (deviation ≤ 7 days), then the causal relationship is determined to be valid.

[0051] Citation relationship verification: For example, extract the document citation entity "Decision on Responsibility Determination" from "Disciplinary Notice (D1)" and match it with the document name of "Decision on Responsibility Determination (B3)". If they match, the citation relationship is determined to be valid.

[0052] In step S223, if a logical relationship is not matched, it is marked as a logical breakpoint, and a repair suggestion is generated based on the rule base. For example, if the "Responsibility Determination Decision (B3)" does not cite the "Job Responsibility Description (B1)", the system will automatically mark it as a "logical breakpoint" and generate a repair suggestion based on the rule base, such as "It is recommended to supplement B3 with the citation of Article 3.2 of the Job Responsibility Clause of B1".

[0053] Finally, a "Logical Relationship Verification Chart" is generated, using lines of different colors to represent logical relationships (green: established; red: breakpoint; yellow: to be verified). Users can click on the lines to view the verification basis (e.g., "Causal relationship established: the time interval of A1's 'Equipment Inspection Leak Filling' and A2's 'Calibration Status Blank' overlaps (2023-11-18 to 2023-11-25)").

[0054] S3. Constructing a Visualized Evidence Chain: After successful verification, a dynamic and interactive visual evidence chain graph is generated. Once all verifications are complete, the system automatically constructs and generates a dynamic and interactive visual evidence chain graph, intuitively displaying the logical network between pieces of evidence. Finally, the complete zeroing case and its structured evidence chain are stored in the knowledge base, supporting semantic-based intelligent retrieval and case recommendation.

[0055] On the other hand, the present invention provides a zeroing system for a zeroing method for managing quality problems in aviation products based on a structured evidence chain, which includes a structured rule knowledge base construction module, an automated verification module, and a visual evidence chain construction module.

[0056] The structured rule knowledge base construction module is used to break down the five-step management zeroing process into multiple secondary sub-steps, define the necessary evidence items and types for each secondary sub-step, form a mapping table of steps, evidence items and types, clearly define the three-element attribute requirements for each evidence item and define the logical relationship rules between different evidence items, thus obtaining the structured rule knowledge base.

[0057] The automated verification module is used for preprocessing evidence and performing triple automated verification.

[0058] The visual evidence chain building module is used to generate a dynamic and interactive visual graph of the evidence chain after verification.

[0059] Specific Implementation Example: This embodiment takes the management and reset of a black screen fault on a certain type of aircraft display screen as an example to further illustrate the method of the present invention.

[0060] Case Background: During operation, multiple aircraft of a certain type experienced an intermittent blackout of the in-cabin system display screen. Technical investigation revealed that the root cause was a cold solder joint issue at specific solder points in the display driver module of batch 05-1. After identifying the technical problem, the company initiated a management-level investigation to systematically determine the management process loopholes and personnel responsibilities that led to this batch of defective products passing inspection and being installed in aircraft.

[0061] Detailed description of the entire system implementation process: S1. System startup and rule base construction.

[0062] Quality Supervisor "Engineer Li" creates a management zeroing task in the system. The system first automatically loads a pre-built structured rule knowledge base according to step S1. This knowledge base defines the five stages of management zeroing → 23 secondary sub-stages → evidence items and ternary attributes, as well as 12 categories of logical relationship rules (58 detailed rules), according to Tables 1 and 2. The system then... Figure 3 Flowchart and Figure 4 The rule base structure has begun to operate.

[0063] S2. Preprocess and perform triple automated verification on the evidence submitted in each sub-step of the zeroing process. This embodiment requires evidence submission and automated verification in five steps. The five steps are as follows: Step 1: Investigate the process by which the problem occurred.

[0064] System boot: The interface prominently displays a prompt that the core of this step is "restoring the scene of management process failure," requiring the submission of two types of evidence: 1) process audit records (examining the system and process itself), and 2) process execution records (reflecting actual operations). The system also displays a pop-up window with regulatory requirements, such as "Audit records must use the company's audit department's standard template, including findings, risk levels, and the auditor's signature."

[0065] User-submitted evidence: Evidence A1 (Documentary Evidence - Audit Record): "Special Audit Report on Welding and Inspection Processes in the Electrical Assembly Workshop (Q4 2023)". The report points out: "Issue P-03: The 'Daily Equipment Inspection and Calibration Record Sheet' for the post-weld optical automatic inspection (AOI) station has omissions and forged signatures, indicating loopholes in process control." Evidence A2 (Documentary Evidence - Execution Record): The "AOI Inspection Log of the Third Batch of DDM-07C Modules" exported from the Manufacturing Execution System (MES). The log shows that for 80 products with module serial numbers DDM-07C-030521 to 030600, the "Equipment Calibration Status" field in the record is either empty or "N / A".

[0066] Evidence A3 (Physical Evidence): High-resolution microscopic photographs and X-ray images of the faulty component. The photographs clearly show a typical "head-in-pillow" solder joint morphology at pin 17 of the U5 chip.

[0067] Evidence A4 (witness testimony): Interview transcript with AOI operator Liu Moumou. In the transcript, Liu Moumou stated: "Around November 20, 2023, the calibration indicator light on my AOI equipment was flashing, but the foreman said that the production task was tight and told me to record 'normal' and continue working, so I did as he said." The system extracts the text content of A1 and A2 using Apache Tika; performs sharpness detection (variance ≥ 100) and character recognition on A3 using OpenCV and OCR; and converts the audio of A4 (if any) to a standard format using FFmpeg and generates a transcript.

[0068] Evidence is subject to triple automated verification, which includes verification of standardization, integrity, and logical relationships.

[0069] Compliance verification: A1 document number and auditor's signature conform to the template; A2 is a structured data table; A3 is a high-resolution image; A4 includes the interviewee's signature and date. All passed.

[0070] Completeness verification: The core documentary evidence categories of "Audit" and "Execution," as well as supplementary "Physical Evidence" and "Testimonial Evidence," have all been submitted. Passed.

[0071] Logical relationship verification: Causal verification: The system's NLP engine extracts the problem description "missing information in inspection record" from A1 and matches it with the specific blank record field in A2. This confirms that A2 is a specific instance of the problem in A1, and the relationship verification passes.

[0072] Spatiotemporal correlation verification: The system identifies "November 20, 2023" mentioned in A4 and compares it with the generation timestamps of 80 abnormal records in A2 (concentrated between 2023-11-18 and 2023-11-25). The time intervals highly overlap, and the logical correlation is enhanced.

[0073] Closed-loop verification of people / objects / documentary evidence: The system establishes a temporary association group: Defect phenomenon (A3 photo of cold solder joint) ← Steps to be detected (A2 AOI inspection log) ← Reasons for failure of this step (A4 Human error) ← Exposed process loopholes (A1 audit report). The system indicates that the logical chain is complete, initially forming an evidence closed loop of "phenomenon-operation-management".

[0074] Step Two: Ascertaining Responsibility.

[0075] System guidance: The goal of this step is to "associate management vulnerabilities with specific positions and personnel", and you need to submit the policy basis, identification documents and personnel confirmation materials.

[0076] User-submitted evidence: Evidence B1 (System Basis): "Job Description for Electrical Assembly Workshop", which clearly stipulates that "AOI Operator (Position): responsible for daily inspection and calibration of equipment and accurate recording, and has the responsibility to stop and report any abnormal equipment status"; "Production Team Leader (Position): responsible for supervising the implementation of process discipline in the team, and has the responsibility to correct and report any violations by operators in a timely manner".

[0077] Evidence B2 (Institutional Basis): "Regulations on Accountability for the Quality of Aviation Products of a Certain Unit".

[0078] Evidence B3 (Document of Determination): "Decision on the Management Responsibility for the Quality Problem of 'Black Screen on a Certain Type of Machine'". Determination: Operator Liu was directly responsible for concealing the equipment malfunction and falsifying records; Team Leader Wang was primarily responsible for violating regulations in command and dereliction of duty in supervision; the Electrical Assembly Workshop was held responsible for inadequate supervision and spot checks on the implementation of the "Equipment Inspection System".

[0079] Evidence B4 (witness confirmation): Receipt of the "Notification of Liability Determination" signed and confirmed by Liu Moumou and Wang Moumou.

[0080] System logical relationship verification: Policy Reference Verification: The system checks whether document B3 explicitly references the specific job responsibility clauses in document B1 (e.g., "Based on Article 3.2 of the Job Responsibility Statement...") and the document number in document B2. Verification passed.

[0081] Responsibility matching verification: The system extracts the responsible persons "Liu Moumou (operator)" and "Wang Moumou (team leader)" identified in B3 and matches them with the positions "AOI operator" and "production team leader" in B1 to confirm that the correspondence between positions and personnel is correct.

[0082] Closed-loop verification with previous steps: The system automatically links the "responsible person: Wang Moumou (illegible command)" in this step with the A4 evidence "the foreman said the production task was urgent..." in the previous step to confirm that the determination of responsibility is supported by factual evidence.

[0083] Step 3: Develop and implement management improvement measures.

[0084] System guidance: The system prompts that measures must be "targeted at the root cause, implementable, and verifiable," and requires the submission of a measure plan, implementation records, and verification results.

[0085] User-submitted evidence: Evidence C1 (Improvement Plan): "Implementation Plan for Strengthening the Management of AOI Equipment Use", the core measures include: ① Activating an electronic inspection system with biometric identification to prevent proxy signing; ② Automatically locking the equipment if the inspection is not completed; ③ Increasing the daily random checks of inspection records by the Quality Department.

[0086] Evidence C2 (Implementation Record): The "Measures Implementation Ledger" shows that the electronic inspection system was launched on January 15, 2024; and the relevant operators and team leaders completed their training on January 10, 2024.

[0087] Evidence C3 (Verification Results): "Verification Report on the Effectiveness of Measures", with screenshots of electronic inspection records from January 16 to February 15, 2024, showing a completion rate of 100%; the Quality Department conducted 10 spot checks and found no abnormalities.

[0088] System logical relationship verification: Problem-Solution Targeting Verification: The system analyzes the measures in C1 (such as "preventing record falsification") to see if they directly address the issues exposed in A1 (audit findings of "omissions and forged signatures") and A4 (verification of "records as 'normal'"). Targeting match successful.

[0089] Closed-loop verification of measures: Check whether the chain of evidence from plan (C1) → implementation (C2) → verification (C3) is complete. The system prompts that the time in the C3 verification report should be after the implementation of C2, and the timing logic is correct.

[0090] Step Four: Serious Handling.

[0091] System guidance: Emphasizing that the handling must be "legal and compliant, and have a warning effect," and requiring the submission of the handling decision, educational materials, and process records.

[0092] User-submitted evidence: Evidence D1 (Handling Decision): The "Notice on Disciplinary Actions Against Comrades Liu and Wang for Violations of Regulations" issued by the company, which imposed corresponding disciplinary action and economic penalties.

[0093] Evidence D2 (Warning Education Material): A courseware titled "Quality Red Line Warning Education Case" compiled based on this incident, including fault pictures, cause analysis, accountability, and lessons learned.

[0094] Evidence D3 (Education Process Record): Sign-in sheet, on-site photos, and training assessment results of the all-staff warning education meeting in the electrical assembly workshop.

[0095] System logical relationship verification: Verification of the basis for processing: Check whether D1 cited B3 (the decision on liability determination) as the basis for processing. Verification passed.

[0096] Education and Handling Collaborative Verification: Check whether the content of the D2 courseware includes the handling results of D1, forming a warning combination of "notification + case analysis". Verification passed.

[0097] Coverage verification: Compare the personnel list on the D3 sign-in sheet with the personnel involved in the positions on B1 (AOI operators, team leaders, etc.) to confirm that all key personnel participated.

[0098] Step 5: Improve regulations.

[0099] System guidance: The goal is to "solidify temporary measures into long-term systems," and a complete set of evidence for system revision, publicity, and implementation inspection must be submitted.

[0100] User-submitted evidence: Evidence E1 (Revision Process): "Approval Form for the Revision of the 'Regulations on the Management of Inspection and Calibration of Production Equipment'", which fully records the entire process from problem identification and departmental countersignature to leadership approval.

[0101] Evidence E2 (New System Text): The officially released "Regulations on the Inspection and Calibration Management of Production Equipment (Rev.3.0)" includes requirements for electronic inspection, abnormal locking, and quality spot checks in the main text.

[0102] Evidence E3 (Dissemination Records): Training plans, courseware, exam papers, and transcripts showing that all participants passed the new regulations.

[0103] Evidence E4 (Implementation Inspection): A special inspection report from the Quality Department within one month after the new regulations took effect, concluding that "the implementation was good and no violations of the new regulations were found."

[0104] System logical relationship verification: Measure-to-system conversion verification: Compare the C1 (temporary measure) and E2 (formal system) texts to confirm that the core requirements have been written into the system.

[0105] Training-Execution Closed-Loop Verification: The confirmation of the E4 inspection report is made after the completion of E3 training, which conforms to the management logic of "training before execution".

[0106] S3. Construct a visual chain of evidence: After successful verification, a dynamic and interactive visual graph of the chain of evidence is generated. A report is also generated and stored in a structured knowledge base.

[0107] Automated construction of visual evidence chains: Once all submitted evidence in the five stages has passed automated verification, the system invokes the evidence chain construction module. This module automatically generates a dynamic, interactive panoramic view of the evidence chain based on the logical relationships established during the verification process (e.g., ...). Figure 5 (Conceptual expansion). In the graph, each piece of evidence is a node, and the connections between nodes represent relationships such as "support," "reference," "cause," and "transformation." When a user clicks on any node (such as the A4 interview transcript), all other evidence that has a direct logical connection to it (such as the A2 detection log that supports it, and the B3 responsibility determination that it caused) is highlighted, making the entire zeroing logic clear at a glance.

[0108] Generate a structured zeroing report: The system report generation module uses a visual evidence chain as its framework, automatically extracting the core summaries, conclusions, and key fields (such as liability determination results and action clauses) from each piece of evidence and populating them into a standardized "Management Zeroing-Out Report" template. Hyperlinks are automatically inserted at key conclusions in the report text, allowing users to trace back to the corresponding original evidence files within the system. Once generated, the report can be directly used for formal zeroing-out review meetings.

[0109] Structured knowledge base storage and intelligent reuse: Storage: This complete zeroing-out case, including all evidence files, evidence chain relationship data, and the final report, was structured, compressed, and stored in the company's quality knowledge base. The system automatically tagged the case with multi-dimensional labels, such as: "Fault location: Display screen / drive module", "Problem type: Welding defect (cold weld)", "Management loophole: Equipment inspection / record falsification", "Involved process: AOI inspection", and "Responsible position: Operator / team leader".

[0110] Learning from experience and reusing solutions: Three months later, when another project at "Company X" encountered a problem with "cracked solder joints in the airborne radar power module," the quality engineer created a new task in the system to resolve the issue. The system's knowledge base module immediately activated its intelligent recommendation system: based on semantic similarity such as "welding defects" and "process control," it automatically pushed the current case of "poor soldering of the display driver module." In the "Investigation Process" section of the newly created task, the system interface sidebar automatically prompted: "Historical similar case suggestions: It is recommended to focus on auditing the equipment inspection and recording standards of the post-welding inspection process. The relevant case evidence chain is ready and can be used for reference." Engineers can import the evidence chain framework of historical cases as a template with one click, greatly accelerating the analysis speed of new problems and truly achieving the management goal of "learning from experience and preventing similar problems."

[0111] The above embodiments fully demonstrate the entire process from S1 rule base construction → S2 evidence preprocessing and triple verification → S3 visualization of the evidence chain generation. The steps are closely connected, verifying the technical feasibility of the present invention.

[0112] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made by those skilled in the art to the technical solutions of the present invention without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A structured chain of evidence based aviation product quality problem management zeroing method, characterized in that: It includes the following steps: S1. The five-step management zeroing process is broken down into multiple second-level sub-steps. The necessary evidence items and their data types are defined for each second-level sub-step. Then, the ternary attribute requirements are defined for each evidence item, and the logical relationship rules between different evidence items are defined and stored in the rule knowledge base. S2. Preprocess and perform triple automated verification on the evidence submitted in each sub-step of the management zeroing process, specifically: S21. Preprocess the evidence and convert multimodal evidence into a structured format; S22. Perform triple automated verification on the evidence. The triple automated verification includes normative verification, integrity verification, and logical relationship verification. The logical relationship verification is specifically as follows: S221. Evidence Entity Extraction: A trained BERT-BiLSTM-CRF model is used to extract core entities from each piece of evidence text. The BERT-BiLSTM-CRF model includes a BERT encoding layer, a BiLSTM temporal feature extraction layer, and a CRF label decoding layer. The BERT-BiLSTM-CRF model employs an improved sequence label loss function, which, based on the standard conditional random field loss, integrates entity boundary penalty terms and label transition constraint terms to enhance its ability to model the boundaries of zero-valued evidence entities and label dependencies. The improved sequence label loss function is as follows: ； where S() is a loss function, N is the number of training samples, is a sample, is a real label sequence of the i-th sample, is all possible label sequences; S222. Construct an entity association graph: Construct an entity association graph based on entity similarity and contextual association; S223. Verify Logical Relationships: A matching algorithm based on graph isomorphism and rule embedding is used to match the entity association graph with the logical relationship rules in the rule knowledge base. Specifically, each logical relationship rule is converted into a graph pattern, where nodes are entity types and edges are relationship types. For the entity association graph G and the rule pattern P, the VF2 subgraph isomorphism algorithm is used to calculate whether there is a subgraph in G that is isomorphic to P. If so, the match is successful; otherwise, the maximum common subgraph is calculated, and the matching degree = |common subgraph| / |P| is output. A first matching degree threshold and a second matching degree threshold are preset and increased sequentially. When the matching degree is greater than or equal to the second matching degree threshold, the logical relationship is considered valid. When the matching degree is greater than or equal to the first matching degree threshold and less than the second matching degree threshold, the logical relationship is considered partially matched. When the matching degree is less than the first matching degree threshold, a logical breakpoint is identified. S3. Construct a visual evidence chain: After verification, a dynamic and interactive visual evidence chain graph is generated.

2. The structured chain-of-evidence based aviation product quality issue management nullification method of claim 1, wherein: Step S1 specifically includes the following sub-steps: S11. The five-step management zeroing process is broken down into 23 secondary sub-steps according to the management zeroing standard process, and each secondary sub-step corresponds to at least one management action. S12. Define the necessary evidence items and their data types for each secondary sub-stage, forming a stage-evidence item-type mapping table. Data types include document type, image type, structured data type, and audio type. S13. Define a ternary attribute requirement for each piece of evidence. The ternary attribute includes normative requirements, completeness requirements, and validity requirements. S14. First-order predicate logic is used to define the logical relationship rules between different evidence items. The logical relationship rules include supporting relationships, referencing relationships, causal relationships and corroborating relationships, and are stored in the rule knowledge base.

3. The structured chain-of-evidence based aerospace product quality issue management nullification method of claim 1, wherein: In step S14, based on the characteristics of the zero-chain evidence, each logical relationship is further subdivided into several subcategories, totaling 12 core logical relationships. Specifically, supporting relationships are divided into direct support, indirect support, and evidence chain support; citation relationships are divided into document citation, clause citation, and external standard citation; causal relationships are divided into direct causation, indirect causation, and conditional causation; and corroboration relationships are divided into witness testimony-documentary evidence corroboration, physical evidence-record corroboration, and time corroboration. Furthermore, 58 subdivided rules are obtained and stored in the relational rule base of the rule engine. The relational rule base is a component of the structured rule knowledge base.

4. The zeroing-out method for aviation product quality problem management based on structured evidence chain according to claim 1, characterized in that: In step S21, the evidence preprocessing specifically involves: extracting text content and metadata for document-type evidence using Apache Tika; performing clarity detection and text recognition for image-type evidence using OpenCV; verifying field integrity for structured data-type evidence using the Python Pandas library; and converting audio-type evidence to a standard format using FFmpeg, while simultaneously generating a text transcript.

5. The zeroing-out method for aviation product quality problem management based on structured evidence chain according to claim 1, characterized in that: In step S22, the normative verification is based on the normative attribute requirements in the rule knowledge base. Metadata verification and content verification are used to perform a two-dimensional verification of normativity and output the verification results. Integrity verification is based on integrity attribute requirements in the rule knowledge base. It uses the integrity of evidence items and the integrity of attachments to perform integrity verification and outputs the verification results.

6. The structured chain-of-evidence based aerospace product quality issue management nullification method according to claim 1, wherein: In step S221, the BERT encoding layer uses a pre-trained BERT model to encode the input evidence text sequence, obtaining the context-related vector representation of each character. The input text sequence is... ,in Let t be the t-th character, and T be the sequence length. After passing through the BERT encoder, the output is the hidden layer vector of each character. , dimension , For the BERT hidden layer dimension, the specific expression is: ； wherein, are trainable parameters of the BERT model.

7. The structured chain-of-evidence based aerospace product quality issue management nullification method of claim 5, wherein: In step S221, the BiLSTM temporal feature extraction layer extracts the context vector output by BERT. The input is a Bidirectional Long Short-Term Memory (BiLSTM) network to capture the forward and backward temporal dependencies of a text sequence. The BiLSTM network consists of a forward LSTM and a backward LSTM. The forward LSTM processes the sequence from left to right and outputs the forward hidden layer vector. The inverse LSTM processes the sequence from right to left and outputs the inverse hidden layer vector. The two are concatenated to obtain the output vector of BiLSTM. , dimension Forward LSTM computation: ； ； ； ； The reverse LSTM computation is symmetrical to the forward one, obtained by concatenating the forward and reverse hidden layer vectors. ,in The activation vectors for the input gate, forget gate, and output gate, respectively. Let be the cell state vector. This is the weight matrix. For bias vectors, For activation function, It is an element-wise product.

8. The structured chain-of-evidence based aerospace product quality issue management nullification method of claim 7, wherein: In step S221, the CRF tag decoding layer introduces a Conditional Random Field (CRF) for tag decoding. The CRF layer outputs a BiLSTM signal. As input, it is first mapped to a label score matrix through a fully connected layer. , dimension , This represents the score for the t-th character labeled as the k-th class label: ； That This is the weight matrix of the fully connected CRF layer. As the bias vector, the goal of the CRF layer is to find the optimal label order. To maximize the sequence score, the sequence score is calculated using the following formula: ； in, This represents the transition weight from label i to label j. The transition matrix, As the starting tag, To terminate the labeling process, reasonable transition relationships between labels are learned through training; the Viterbi algorithm is used to solve for the optimal label sequence. : 。 9. The structured chain-of-evidence based aerospace product quality issue management nullification method according to claim 1, wherein: The specific steps of step S222 are as follows: Entity similarity calculation: Cosine similarity is used to calculate the semantic similarity between text entities. Combined with attribute similarity, candidate entity pairs with similarity higher than the similarity threshold θ_sim are selected. Contextual association mining: Constructing association edges based on the temporal, spatial, and logical attributes of evidence; Graph construction: Using each extracted entity as a node and similarity or association as an edge, an initial entity association graph G(V,E) is generated. Isolated nodes and weak association edges with weights lower than the weight threshold θ_edge are removed to obtain the final entity association graph.

10. A zeroing system for the structured chain of evidence based aviation product quality problem management zeroing method of claim 1, characterized in that: It includes a structured rule knowledge base construction module, an automated verification module, and a visual evidence chain construction module. The structured rule knowledge base construction module is used to break down the five-stage management zeroing process into multiple secondary sub-stages, define the necessary evidence items and types for each secondary sub-stage, form a mapping table of stages, evidence items and types, clearly define the three-element attribute requirements for each evidence item and define the logical relationship rules between different evidence items, thus obtaining the structured rule knowledge base; The automated verification module is used for preprocessing evidence and performing triple automated verification. The visual evidence chain building module is used to generate a dynamic and interactive visual graph of the evidence chain after verification.

Images