Vector matching-based abnormal luggage processing method and system

Abstract

The invention relates to the technical field of abnormal luggage processing, and discloses an abnormal luggage processing method and system based on vector matching, and the method comprises the steps: constructing a mixed retrieval knowledge base, integrating multi-source heterogeneous data to generate a relational type and a vector knowledge base, and carrying out the index association; establishing a mixed retrieval rule engine, performing keyword analysis and semantic vectorization on the case description, generating a query instruction, and performing parallel retrieval to obtain a candidate scheme set; the development field enhancement decision system generates a structured processing suggestion through a multi-dimensional evidence chain and an RAG framework, and visually outputs a process specification and a compensation basis; the disposal system is constructed to integrate all the modules, and full-process automation is achieved. The system comprises a hybrid retrieval knowledge base construction module, a rule engine module, a domain enhancement decision module and an integration module. Vector matching and rule retrieval are combined, decision making is enhanced through a large model, the abnormal luggage processing efficiency, normalization and scientificity are improved, and the method is suitable for intelligent abnormal luggage processing in air transportation.

Description

Technical Field

[0001] The present invention relates to the technical field of abnormal luggage handling, and specifically to an abnormal luggage handling method and system based on vector matching. Background Art

[0002] In the field of air transportation, the handling of abnormal luggage is an important link in protecting the rights and interests of passengers and the quality of air services. With the increasing demand for air travel, abnormal luggage cases have shown a trend of diversification and complexity, and traditional handling methods face many challenges.

[0003] From the perspective of data processing, abnormal luggage cases involve multi-source heterogeneous data, including policy documents of different airlines, historical case records, luggage damage images, etc. Traditional methods lack effective integration and utilization of these data: on the one hand, paper-based or fragmented policy documents are difficult to quickly retrieve and structurally analyze, resulting in insufficient consistency and traceability of processing rules; on the other hand, historical case data has not been fully mined for semantic associations, and case similarity analysis cannot provide references for new cases, resulting in high costs for repetitive labor and low processing efficiency.

[0004] In the retrieval and decision-making link, the traditional processing process relies on manual access to rules and empirical judgments, which has significant limitations. Manual keyword retrieval is difficult to capture deep semantic associations. For example, the definitions and compensation standards for "luggage damage" may vary among different airlines, and simple keyword matching cannot accurately adapt to specific scenarios; at the same time, manual decision-making lacks multi-dimensional evidence support, and liability determination and compensation calculation are easily affected by subjective factors, resulting in insufficient fairness and transparency of processing results and a high passenger complaint rate.

[0005] In terms of visualization and interaction, traditional systems lack an intuitive process display and a human-machine collaboration mechanism. The processing process is usually presented in the form of a text report, and the key decision-making nodes and regulatory bases are not clear, making it difficult for front-line staff to quickly master the processing logic; in addition, the lack of visual case comparison and risk assessment tools results in the lack of data support for the optimization and adjustment of processing plans and makes it difficult to adapt to the dynamic change requirements of complex cases.

[0006] With the development of artificial intelligence technology, technologies such as vector matching and large language models provide new paths to solve the above problems. However, the existing technologies have not yet formed a complete system for handling abnormal luggage: on the one hand, the fusion mechanism of vector retrieval and rule retrieval is not mature enough to achieve efficient collaborative query of cross-modal data; on the other hand, the depth of domain knowledge enhancement and decision-making reasoning is insufficient, making it difficult to organically combine policy rules, historical cases, and real-time case data, resulting in limited accuracy and reliability of automated processing. Therefore, there is an urgent need for a method and system for handling abnormal luggage that can integrate multi-source data, achieve intelligent retrieval and decision-making, so as to improve processing efficiency, standardize processing processes, and ensure the scientificity and fairness of processing results. Summary of the Invention

[0007] The purpose of the present invention is to provide a method and system for handling abnormal luggage based on vector matching to solve the problems raised in the above background technology.

[0008] To achieve the above purpose, the present invention provides the following technical solution: A method for handling abnormal luggage based on vector matching, the method includes: Step S1: Construct a hybrid retrieval knowledge base; integrate multi-source heterogeneous data of abnormal luggage case data to generate a relational knowledge base and a vector knowledge base; index and associate the relational knowledge base with the vector knowledge base to establish a hybrid retrieval knowledge base; Step S2: Establish a hybrid retrieval rule engine; perform keyword parsing and semantic vectorization on the input description of abnormal luggage cases to generate a hybrid retrieval query instruction; perform parallel retrieval based on the hybrid retrieval knowledge base to generate a set of candidate processing solutions; Step S3: Develop a domain-enhanced decision-making system based on a large model; enhance the domain knowledge of the set of candidate processing solutions to generate structured processing suggestions; output the processing process specifications and compensation basis through a visual interaction interface; Step S4: Construct an abnormal luggage disposal system; integrate the hybrid retrieval rule engine, the domain-enhanced decision-making system, and the visual interaction interface to achieve full-process automated processing of abnormal luggage cases from retrieval to decision-making.

[0009] Preferably, step S1 includes the following steps: Step S11: Construct a relational knowledge base; collect abnormal luggage handling policy documents of different airlines, perform document structuring processing through a localization parsing tool to generate an airline-abnormal type-processing rule triple data set; Step S12: Construct a vector knowledge base; collect historical abnormal luggage case data, perform missing value cleaning and field standardization processing, generate text vectors and multi-modal joint vectors through a pre-trained language model, and store them in a vector index library; Step S13: Establish an index association mechanism; set the airline code primary key for the relational knowledge base, configure a dynamic sharding index for the vector knowledge base, and establish a cross-modal query channel for the hybrid retrieval knowledge base through primary key mapping.

[0010] Preferably, step S2 includes the following steps: Step S21: Parse the input case description, extract the airline code, anomaly type, and item feature fields, and generate a keyword retrieval expression; Step S22: Perform semantic encoding on the unstructured text to generate a query vector; calculate the cosine similarity between the query vector and the vectors in the vector knowledge base, and screen the top N similar cases according to the score ranking; Step S23: Merge the keyword retrieval results and the vector retrieval results, perform conflict resolution processing, and generate a candidate processing solution set including the processing flow, compensation standard, and historical cases.

[0011] Preferably, step S22 includes the following steps: Step S221: Segment and perform entity recognition on the case description text, and extract the airline identifier, abnormal item category, and damage level features; Step S222: Use a pre-trained language model to generate text semantic vectors, and perform feature dimensionality reduction through a multi-layer perceptron to generate a query vector with a fixed dimension; Step S223: Perform sharded parallel retrieval in the vector index library, calculate the similarity score using the weighted cosine similarity algorithm, and take the top N similar cases after merging the sharded results.

[0012] Preferably, step S3 includes the following steps: Step S31: Perform domain knowledge enhancement; extract the airline-specific processing rules from the relational knowledge base, retrieve the similar case processing records from the vector knowledge base, and form a multi-dimensional evidence chain; Step S32: Construct a RAG inference framework; convert the retrieval results into structured prompt words, and input them into the large language model to generate processing suggestions, including the liability determination logic and the basis for calculating the compensation amount; Step S33: Design a visual output template; convert the processing suggestions into a structured flow chart, mark the key decision nodes and the source of regulatory references, and generate an editable disposal plan document.

[0013] Preferably, step S31 includes the following steps: Step S311: Establish a rule weight matrix; assign weight coefficients to different sources of evidence according to the airline policy priority and the confidence of processing cases; Step S312: Perform evidence fusion calculation; use a Bayesian network model to integrate structured rules and unstructured case data, and generate a comprehensive processing suggestion confidence score; Step S313: Build a decision tree model; convert the fusion result into a decision tree structure including abnormal type judgment, compensation standard application, and processing flow path.

[0014] Preferably, step S4 includes the following steps: Step S41: Deploy a hybrid retrieval engine module; configure a multi-threaded concurrent controller to achieve synchronous query scheduling of relational databases and vector index libraries; Step S42: Develop a decision inference engine; integrate a rule engine and a probabilistic inference model to support compliance verification and risk assessment of processing solutions; Step S43: Build a human-computer interaction layer; design a multi-view visualization panel to support functions such as comparison and analysis of processing solutions, traceability of regulatory clauses, and electronic signature confirmation.

[0015] Preferably, in the construction of the vector knowledge base in step S12, an adversarial training mechanism is introduced, and adversarial luggage damage image data is generated through a GAN network.

[0016] Preferably, the domain-enhanced decision system in step S3 includes a dynamic knowledge distillation module: establish a three-layer architecture including a basic rule layer, a case reasoning layer, and a meta-learning layer, and gradually increase the training weights of complex cases through a curriculum learning strategy.

[0017] Preferably, the present invention further includes an abnormal luggage processing system based on vector matching, and the system includes: A hybrid retrieval knowledge base construction module for integrating multi-source heterogeneous data of abnormal luggage case data to generate a relational knowledge base and a vector knowledge base, and associating the two through indexing to form a hybrid retrieval knowledge base; A hybrid retrieval rule engine module for parsing keywords and semantic vectorization of the input abnormal luggage case description to generate a hybrid retrieval query instruction, and performing parallel retrieval based on the hybrid retrieval knowledge base to output a candidate processing solution set; A domain-enhanced decision module integrating a large language model for enhancing domain knowledge of the candidate processing solution set to generate a structured processing suggestion, and outputting a processing flow specification and compensation basis through a visual interaction interface; An integration module for logically integrating the hybrid retrieval knowledge base construction module, the hybrid retrieval rule engine module, and the domain-enhanced decision module to achieve full-process automated processing of abnormal luggage cases from retrieval to decision-making.

[0018] Compared with the prior art, the beneficial effects of the present invention are: In terms of data integration and management, relational knowledge bases and vector knowledge bases are generated through the integration of multi-source heterogeneous data, and an index association mechanism is established. The relational knowledge base structures the policy documents of different airlines into "airline - exception type - processing rule" triples, solving the problems of fragmentation and difficult retrieval of traditional policy documents, and ensuring the consistency and traceability of processing rules; the vector knowledge base generates text vectors and multi-modal joint vectors through pre-trained language models, and at the same time introduces an adversarial training mechanism to generate adversarial luggage damage image data, not only realizing the semantic-level representation of historical case data, but also enhancing the model's generalization ability for complex damage scenarios. The index association between the two establishes a cross-modal query channel, providing a data basis for subsequent hybrid retrieval and improving the depth and breadth of data utilization.

[0019] In the intelligent retrieval process, the hybrid retrieval rule engine generates hybrid retrieval query instructions through keyword parsing and semantic vectorization, realizing the parallel execution of rule retrieval and vector retrieval. Keyword retrieval accurately extracts structured information such as airline codes and exception types, quickly locating and matching policy rules; semantic vectorization generates query vectors through word segmentation, entity recognition, and pre-trained models, calculating cosine similarity to screen similar cases and capturing the deep semantic associations of unstructured text. After merging the results of the two, conflict resolution is performed to generate a candidate solution set containing processing procedures, compensation standards, and historical cases. Compared with traditional single retrieval methods, it significantly improves the comprehensiveness and accuracy of retrieval, and shortens the response time for case processing.

[0020] At the decision support level, the domain-enhanced decision-making system realizes the intelligent generation of processing suggestions through multi-dimensional evidence chain construction and the RAG reasoning framework. Extract airline-specific rules from the relational knowledge base, retrieve similar case processing records from the vector knowledge base, form an evidence chain combining rules and cases, and perform evidence fusion through a rule weight matrix and a Bayesian network model to generate a comprehensive confidence score, ensuring the scientific nature of decision-making basis. The RAG framework converts the retrieval results into structured prompt words and inputs them into a large language model to generate processing suggestions containing liability determination logic and compensation calculation basis. At the same time, the fusion results are structured through a decision tree model to clarify key nodes such as exception type judgment and compensation standard application. The three-layer architecture of the dynamic knowledge distillation module (basic rule layer, case reasoning layer, meta-learning layer) gradually increases the weights of complex cases through a curriculum learning strategy, continuously improving the model's ability to handle complex cases, making the decision results more in line with actual business needs, reducing the cost of manual intervention, and improving the fairness and transparency of processing results.

[0021] In terms of system integration and interaction, the abnormal baggage disposal system realizes the full-process automation from retrieval to decision-making. The multi-threaded concurrent controller of the hybrid retrieval engine module ensures the synchronous scheduling of the relational database and the vector index library, improving the retrieval efficiency; the decision-making inference engine integrates the rule engine and the probability inference model, supports compliance verification and risk assessment, and guarantees the legality and security of the processing plan; the multi-view visualization panel of the human-computer interaction layer supports plan comparison, regulation traceability and electronic signature confirmation, making the processing process intuitive and transparent, facilitating the operation of front-line personnel and communication with passengers, enhancing the user experience and reducing disputes caused by unclear processes.

[0022] Overall, through technology integration and system innovation, the present invention constructs an intelligent and standardized abnormal baggage processing system, effectively solving problems such as insufficient data utilization, low retrieval and decision-making efficiency, and strong subjectivity of processing results in traditional processing methods, significantly improving the efficiency, standardization and scientificity of abnormal baggage processing, and providing strong technical support for service optimization in the air transportation industry and protection of passengers' rights and interests. Brief Description of the Drawings

[0023] Figure 1 It is the working principle diagram of the abnormal baggage processing method based on vector matching described in the present invention; Figure 2 It is the design diagram for constructing a hybrid retrieval knowledge base; Figure 3 It is the design diagram for establishing a hybrid retrieval rule engine; Figure 4 It is the design diagram for developing a domain-enhanced decision-making system based on a large model; Figure 5 It is the design diagram for constructing an abnormal baggage disposal system. Detailed Embodiments

[0024] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0025] Please refer to Figures 1 - 5 , the abnormal baggage processing method and system based on vector matching involved in the present invention are specifically implemented as follows: Step S1: Construct a hybrid retrieval knowledge base; perform multi-source heterogeneous data integration on abnormal baggage case data to generate a relational knowledge base and a vector knowledge base, and index and associate the relational knowledge base with the vector knowledge base to establish a hybrid retrieval knowledge base.

[0026] Step S2: Establish a hybrid retrieval rule engine; perform keyword parsing and semantic vectorization on the input description of abnormal luggage cases to generate hybrid retrieval query instructions; execute parallel retrieval based on the hybrid retrieval knowledge base to generate a set of candidate processing solutions.

[0027] Step S3: Develop a domain-enhanced decision-making system based on a large model; enhance the domain knowledge of the set of candidate processing solutions to generate structured processing suggestions; output the processing flow specifications and compensation basis through a visual interactive interface.

[0028] Step S4: Build an abnormal luggage disposal system; integrate the hybrid retrieval rule engine, the domain-enhanced decision-making system, and the visual interactive interface to achieve full-process automated processing of abnormal luggage cases from retrieval to decision-making.

[0029] The present invention will be further described below in conjunction with Embodiments 1 to 4: Embodiment 1: In the process of constructing the hybrid retrieval knowledge base in Step S1, the specific implementation method is as follows: Step S11 constructs a relational knowledge base. Collect abnormal luggage handling policy documents of different airlines, and these document sources include but are not limited to the luggage transportation rule manuals officially released by each airline, contract agreements related to luggage handling, industry specification documents, etc. Since the policy documents of different airlines may exist in different formats, such as PDF documents, Word documents, picture-format announcements, etc., it is necessary to perform structured processing on these documents through a localization parsing tool. The localization parsing tool can adopt corresponding parsing algorithms according to different document formats. For example, for PDF documents, use OCR optical character recognition technology to extract text content; for structured table data, directly read the field information in the table. During the parsing process, it is necessary to identify the key information in the policy documents, including airline names or codes, abnormal luggage types (such as luggage loss, damage, delay, damage to contents, overweight, etc.), and corresponding handling rules (such as compensation standards, processing flow timeliness, liability determination basis, etc.). Through the extraction and structured processing of the key information, a triple dataset of airline-abnormal type-handling rule is generated. For example, for an airline with the code CA, the abnormal type is luggage damage, and the handling rule may include specific terms such as "Verify the damage situation within 72 hours. If it is caused by the carrier, compensate no more than 100 yuan per kilogram of luggage, with a maximum of no more than 2000 yuan". Store these triple data according to the table structure of the relational database to form a relational knowledge base, where each triple is used as a record, and the airline code and abnormal type are used as the primary key fields, and the handling rule is used as the detailed description field, which is convenient for quickly locating the abnormal luggage handling policy of a specific airline through keyword retrieval.

[0030] Step S12 constructs a vector knowledge base. Historical abnormal luggage case data is collected, and the data sources include luggage handling work orders recorded within the airline, case descriptions in the customer service system, filing information at the airport luggage inquiry desk, etc. These data may have problems such as inconsistent formats, missing fields, and ambiguous expressions, so data preprocessing needs to be carried out first. First, missing value cleaning is performed. For records missing key fields such as airline code, abnormal type, case occurrence time, item characteristics, etc., they are completed by manual verification or comparison with data from other associated systems. If they cannot be completed, they are excluded. Then, field standardization processing is carried out to unify fields with the same meaning but different expressions into a standard format. For example, "luggage lost" and "lost luggage" are unified as "lost", and "laptop computer" and "computer" in item characteristics are unified as "laptop computer", etc. The preprocessed case data includes structured data (such as airline code, abnormal type, processing result, etc.) and unstructured data (such as case detailed description text, luggage damage photos, surveillance video clips, etc.). For text data, vectorization processing is carried out through pre-trained language models (such as BERT, RoBERTa, etc.) to convert the text into dense vectors in a high-dimensional vector space to capture the semantic information in the text. For multimodal data (such as luggage damage images), multimodal pre-trained models (such as CLIP, ViT-BERT, etc.) are used to generate a joint representation of text vectors and image vectors to form multimodal joint vectors. All generated vectors are stored in the vector index library, and the vector index library adopts an efficient index structure (such as FAISS, Annoy, etc.) for subsequent rapid retrieval. It should be noted that an adversarial training mechanism is introduced during the construction of the vector knowledge base, and adversarial luggage damage image data is generated through a GAN network (Generative Adversarial Network). Specifically, the generator network learns the distribution characteristics of real luggage damage images and generates adversarial samples that are highly similar to the real images but have slight differences. The discriminator network is responsible for distinguishing between real images and generated images. Through the adversarial training of the two, the feature extraction ability of the model for complex and variable luggage damage images is improved, and the robustness and generalization ability of the vector knowledge base are enhanced. The generated adversarial image data is mixed with the real data and stored in the vector index library after manual review to enrich the sample diversity of the vector knowledge base.

[0031] Step S13 establishes an index association mechanism. Set the airline code primary key for the relational knowledge base. The airline code uses the internationally common two-letter code (e.g., CA represents Air China, MU represents China Eastern Airlines, etc.) to ensure the uniqueness and standardization of the primary key. For the vector knowledge base, configure a dynamic sharding index. Divide the vector data into multiple shards according to dimensions such as the feature distribution or timestamp of the data, and establish an index for each shard independently to support parallel retrieval and improve retrieval efficiency. Establish a cross-modal query channel for the hybrid retrieval knowledge base through primary key mapping. The specific implementation method is as follows: Retain the airline code field in both the relational knowledge base and the vector knowledge base. When a cross-modal query is required, first locate the processing policy of a specific airline through the airline code primary key in the relational knowledge base, and then retrieve the corresponding historical case vector data in the vector knowledge base according to the airline code to achieve the associated retrieval of structured rules and unstructured cases. For example, when querying the luggage loss handling plan of a certain airline (code CZ), first obtain the handling rules of this airline regarding luggage loss through the relational knowledge base, and then retrieve the similar historical case vectors in the vector knowledge base through the airline code CZ to obtain the handling records of similar cases, thus forming a hybrid retrieval result containing rules and cases. This index association mechanism breaks the barrier between structured data and unstructured data, enabling the system to utilize both the accuracy of rules and the experience of cases simultaneously, providing a solid foundation for subsequent hybrid retrieval.

[0032] During the construction process of the entire hybrid retrieval knowledge base, it is necessary to ensure the accuracy and consistency of the data. For the processing rules in the relational knowledge base, regularly compare them with the latest policy documents issued by each airline and update the outdated or invalid rules in a timely manner; for the historical case data in the vector knowledge base, regularly perform data cleaning and deduplication processing, delete duplicate records or invalid data, and ensure the quality of the vector index library. At the same time, establish a data security protection mechanism to encrypt the storage and access control of sensitive data (such as passenger personal information, details of the airline's unpublicized handling policies, etc.) to prevent data leakage or abuse. Through the above steps, a hybrid retrieval knowledge base with clear structure, efficient index, and rich data is constructed, providing strong data support for the retrieval and handling of abnormal luggage cases.

[0033] Embodiment 2: During the process of establishing the hybrid retrieval rule engine in step S2, the specific implementation method is as follows: Execute step S21 to parse the input case description and generate a keyword retrieval expression. The system receives the abnormal luggage case description input by the user, which can be in text form (such as manually entered case details, customer service chat records) or the text content after voice conversion. For the input text, first perform word segmentation. Adopt a dictionary-based word segmentation algorithm (such as forward maximum matching algorithm, reverse maximum matching algorithm) combined with a deep learning word segmentation model (such as LSTM-CRF) to split the continuous text sequence into independent words or phrases. During the word segmentation process, improve the word segmentation accuracy through a custom domain dictionary (including professional vocabulary such as airline codes, abnormal types, item names, etc.). For example, recognize "Air China" as "CA" (airline code), "crack on the suitcase" as "damage" (abnormal type), etc. After word segmentation, perform entity recognition. Use a named entity recognition (NER) model (such as BERT-NER) to extract key entities from the word segmentation results, including airline codes, abnormal types, item feature fields, etc. For example, when the input case description is "A passenger on China Eastern Airlines flight MU5102 reported that the checked luggage was lost, including a laptop and several pieces of clothing", the airline code "MU", abnormal type "lost", and item features "laptop, clothing" can be extracted through entity recognition. According to the extracted entity information, generate a keyword retrieval expression in the format of "airline code = MU AND abnormal type = lost AND item feature = laptop". This expression is used to retrieve the corresponding processing rules in the relational knowledge base, such as the processing process and compensation standards of China Eastern Airlines for lost luggage containing electronic products.

[0034] Execute step S22 to perform semantic vectorization on the unstructured text and retrieve similar cases. First, perform semantic encoding on the unstructured text part (such as the detailed event process, passenger statement, etc.) in the case description. The specific steps are as follows: Step S221: Perform word segmentation and entity recognition on the case description text, which is similar to the word segmentation and entity recognition processes in step S21, but here it focuses more on extracting key features at the semantic level, such as airline identifiers (such as "China Eastern Airlines" corresponding to the code "MU"), abnormal item categories (such as "electronic products", "clothing"), damage level features (such as "slight damage", "severe deformation", "completely lost", etc.). Through entity recognition, convert the features described in natural language into standardized labels that can be recognized by the system. For example, convert "the wheel of the suitcase is broken" to "damage - component damage", and convert "the luggage has not been delivered for more than 72 hours" to "delay - overtime".

[0035] Step S222: Generate text semantic vectors using a pre-trained language model. The pre-trained language model selected is a model fine-tuned on a large amount of text data in the aviation field (such as RoBERTa-wwm-ext-finetuned-airline), which can better capture the semantic associations of aviation field terms. The processed text is input into the model, and the hidden layer vector of a fixed dimension is output as the text semantic vector. To reduce the vector dimension to improve computational efficiency, feature dimensionality reduction is performed through a multi-layer perceptron (MLP), mapping the high-dimensional vector to a query vector of a preset dimension (such as 768 dimensions). For example, for the text describing "the handlebar of the luggage broke during transportation due to improper loading and unloading", the generated query vector needs to contain vector representations of semantic features such as "improper loading and unloading" and "broken handlebar".

[0036] Step S223: Perform sharded parallel retrieval in the vector index library. The vector index library adopts a dynamic sharding mechanism, dividing the vector data into multiple logical shards (such as sharding by airline code, sharding by exception type, etc.), and each shard is assigned an independent retrieval thread. When a query vector is received, the system sends retrieval requests to all shards, and each shard calculates the similarity score between the query vector and the vectors within this shard in parallel. The similarity calculation uses the weighted cosine similarity algorithm, setting weight coefficients for different feature dimensions. For example, the weight of the exception type dimension is higher than that of the item color dimension to highlight the influence of key features on similarity. The weight coefficients are determined through domain expert experience or historical data statistics. For example, the weight of the airline identifier is set to 0.4, the weight of the exception item category is set to 0.3, and the weight of the damage level is set to 0.3. After each shard finishes the calculation, it returns the top K cases with similarity scores. The system merges the results of all shards, sorts them in descending order of score, and selects the top N similar cases (N is a preset threshold, such as 50) as the vector retrieval result. For example, if the weighted cosine similarity between the query vector and a historical case vector is 0.85 and it ranks 3rd, then this case is included in the candidate result set.

[0037] Subsequently, step S23 is executed to merge the retrieval results and generate a candidate processing solution set. The keyword retrieval results of step S21 (processing rules from the relational knowledge base) are fused with the vector retrieval results of step S22 (historical cases from the vector knowledge base). First, a data mapping relationship is established to associate the processing rules in the keyword retrieval results with the historical cases in the vector retrieval results according to airline code and exception type. For example, the baggage loss rules and loss cases of the same airline are associated. Then, conflict resolution processing is performed. When there is an inconsistency between the rule and the case (such as the rule stipulates a compensation ceiling of 2,000 yuan, and the compensation amount in a certain case is 2,500 yuan), it is resolved through the following strategies: preferentially adopt the most recently effective rule. If the case is a historical record before the rule update, the rule shall prevail; if the case is a special processing case after the rule update, extract the special circumstances in the case (such as the passenger purchased baggage insurance, the value of the items in the baggage was notarized, etc.) as a supplementary explanation to the rule. After conflict resolution, information such as the processing flow, compensation standard, and historical cases is integrated to generate a candidate processing solution set. Each solution includes the following content: Airline information: such as airline name, code; Exception type: such as loss, damage; Processing rules: refer to the specific clauses in the relational knowledge base, such as Article X of the "Baggage Transportation Rules of XX Airlines"; Similar cases: list the top M cases (M ≤ N) retrieved by the vector, including case numbers, processing results, and key feature comparisons; Compensation standard: combine the rules and cases to clarify the compensation amount calculation method, time limit requirements, etc.; Processing flow: show the operation specifications of each link in the form of a timeline, such as "file a case within 24 hours → verify within 48 hours → complete compensation approval within 7 working days".

[0038] During the establishment of the entire hybrid retrieval rule engine, it is necessary to optimize the retrieval efficiency and accuracy. For keyword retrieval, accelerate the query by establishing database indexes (such as airline code index, exception type index); for vector retrieval, regularly update the vector index library, delete invalid case vectors, add new case vectors, and recalculate the shard index. At the same time, provide an interface for manual intervention. When the candidate solutions generated by the system do not meet the actual requirements, the operator can manually adjust the retrieval keywords, similarity threshold, or conflict resolution strategy to ensure that the retrieval results meet the business scenario. In addition, adopt a log recording mechanism to record the input parameters, retrieval results, processing time, etc. of each retrieval, which is convenient for subsequent system performance analysis and optimization. Through the above steps, the hybrid retrieval rule engine can efficiently and accurately generate a candidate processing solution set containing multi-dimensional information, providing rich input materials for subsequent domain knowledge enhancement and decision-making.

[0039] Example 3: In the process of developing the domain-enhanced decision-making system based on the large model in step S3, the specific implementation method is as follows: First, perform step S31 for domain knowledge enhancement. Extract the airline-specific processing rules from the relational knowledge base. For example, the compensation rule of a certain airline for "international flight luggage delay exceeding 48 hours" is "compensate at a standard of $50 per day, with a maximum of $300"; at the same time, retrieve the processing records of similar cases from the vector knowledge base. For example, the processing result of "delay for 72 hours and there are urgently needed medicines in the luggage" in historical cases is "priority delivery + additional compensation of $200". Combine the rules and cases to form a multi-dimensional evidence chain including liability determination basis, compensation standard, and processing process.

[0040] To integrate evidence from different sources, establish a rule weight matrix (step S311). According to the airline policy priority (such as IATA standard > airline general rules > local branch rules) and the confidence level of processing cases (based on the historical accuracy of case processing results), assign weight coefficients to each piece of evidence. Let the evidence set be , and the corresponding weight coefficients be , satisfying . Among them, represents the th piece of evidence (such as a certain airline rule or a certain historical case), and represents the credibility weight of this evidence, which is comprehensively evaluated by domain experts according to the authority and timeliness of the evidence.

[0041] Then perform evidence fusion calculation (step S312), and use the Bayesian network model to integrate structured rules and unstructured case data. The nodes of the Bayesian network include abnormal types (such as loss, damage), responsible parties (such as carriers, passengers, third parties), compensation amount intervals, etc., and the edges represent the probabilistic dependence relationships between the nodes. Train the conditional probability table of the Bayesian network through historical data and calculate the confidence score of the comprehensive processing suggestion. Let be the confidence level of the processing suggestion, and the calculation formula is: .

[0042] Among them, represents the conditional probability of the processing result supported by the evidence , which is obtained through Bayesian network reasoning. This formula fuses the support degrees of different evidences for the processing result through weighted summation to generate a comprehensive confidence score, which is used to measure the reliability of the processing suggestion.

[0043] Subsequently, a decision tree model is constructed (step S313), and the fusion result is transformed into a tree structure. The root node of the decision tree is the abnormal type judgment (such as "whether it is luggage damage"), the intermediate nodes are the detailed features (such as "whether the damage degree exceeds 50%" and "whether there are obvious external force marks"), and the leaf nodes are the processing conclusions (such as "compensate at full price" and "compensate after repair"). Each path corresponds to a decision logic chain. For example: "Luggage damage → Damage degree > 50% → Carrier's liability → Compensate according to the current market value".

[0044] In step S32, a RAG inference framework (Retrieval-Augmented Generation framework) is constructed. The evidence chain and decision tree structure extracted in step S31 are converted into structured prompting words in the format of "[Case features] + [Evidence set] + [Decision logic] + Content to be generated". For example, the prompting word is "Case type: International flight luggage loss, airline code: CA, loss time: over 72 hours, evidence: Article X of CA airline regulations (compensation limit of 2,000 yuan), similar case Y (compensation of 1,800 yuan), please generate the liability determination logic and the basis for calculating the compensation amount". The prompting word is input into a large language model (such as the fine-tuned LLaMA-2), and the model generates processing suggestions based on the pre-trained knowledge and the input prompt. The content includes the liability determination logic (such as "According to the Warsaw Convention and CA airline regulations, the carrier fails to prove that it has fulfilled its reasonable care obligation and needs to bear full liability") and the basis for calculating the compensation amount (such as "Combining the rule limit of 2,000 yuan and the similar case of 1,800 yuan, considering the depreciation factor of the item, it is recommended to compensate 1,500 yuan").

[0045] In step S33, a visual output template is designed. The processing suggestions are transformed into a structured flowchart. The swimlane diagram form is used to distinguish the operation processes of different responsible parties (such as airlines, passengers, airports), and key decision nodes (such as "whether to confirm the responsible party" and "whether the compensation amount exceeds 5,000 yuan requires higher-level approval") and the source of regulatory references (such as Article 125 of the Civil Aviation Law) are marked. At the same time, an editable disposal plan document is generated, including a tabular comparison of compensation standards, a case feature comparison table, etc., to facilitate operators to quickly understand and adjust the plan.

[0046] The domain-enhanced decision system includes a dynamic knowledge distillation module and establishes a three-layer architecture: The basic rule layer stores standardized processing rules, the case reasoning layer stores the feature vectors and processing logics of historical cases, and the meta-learning layer automatically adjusts the training weights through a curriculum learning strategy. The curriculum learning strategy gradually increases the training weights according to the case complexity (such as simple loss → complex damage → multi-party liability disputes), enabling the system to first master the basic rules and then learn the processing logics of complex cases, improving the generalization ability and decision-making accuracy of the model.

[0047] Through the above steps, the domain-enhanced decision-making system can generate structured processing suggestions based on multi-source evidence, and clearly present the processing process and compensation basis through a visual interface, providing intelligent support for the automated decision-making of abnormal luggage cases.

[0048] Embodiment 4: During the process of constructing the abnormal luggage disposal system in step S4, the specific implementation method is as follows: Execute step S41 to deploy the hybrid retrieval engine module. This module needs to implement the synchronous query scheduling of a relational database (such as MySQL) and a vector index library (such as FAISS), and improve the retrieval efficiency by configuring a multi-threaded concurrent controller. Taking an abnormal luggage case received by a certain airline as an example, a passenger reported through the customer service system that the checked luggage on flight CA1203 was damaged and contained photographic equipment. After the system received the case description, the multi-threaded concurrent controller started two independent threads: one thread accessed the relational database and retrieved the corresponding processing rules according to the extracted airline code "CA" and abnormal type "damage". For example, Air China stipulates that "for damaged checked luggage, the damage appraisal shall be completed within 48 hours. If it is caused by mechanical failure, it shall be compensated according to the actual loss but not exceeding 3,000 yuan"; the other thread accessed the vector index library, generated a query vector for the case description and retrieved similar cases, and returned the top 50 cases with the highest similarity among 150 luggage damage cases handled by Air China in the past six months. Among them, 30 cases involved photographic equipment, and the case details showed that most of them were compensated according to the amount of the equipment repair invoice. The two threads executed in parallel and completed the data retrieval within about 200 milliseconds, avoiding the problem of excessive time consumption that may be caused by traditional serial retrieval. The multi-threaded controller is also responsible for coordinating between threads. For example, when a thread times out or returns abnormal data, it automatically triggers a retry mechanism or switches to a backup index node to ensure the stability of the retrieval process.

[0049] Execute step S42 to develop a decision-making reasoning engine. The engine integrates a rule engine and a probabilistic reasoning model for compliance verification and risk assessment of treatment plans. Taking the damaged luggage case of flight CA1203 as an example, the candidate treatment plan set obtained from the hybrid search engine includes three types of plans: Plan 1 is based on the basic rules of Air China, with compensation of 100 yuan per kilogram (assuming that the luggage weighs 5 kilograms, compensation is 500 yuan); Plan 2 refers to similar cases and compensates according to the repair costs of photographic equipment (estimated to be 2,000 yuan); Plan 3 combines the latest industry specifications. If the passenger can provide proof of the value of the equipment, compensation can be made according to the actual value (the passenger provided a purchase invoice, which is worth 5,000 yuan). The rule engine first verifies whether each plan complies with current regulations and airline policies. It is found that although Plan 1 complies with the basic rules of Air China, it does not consider the supplementary provisions on the declaration of valuables in the new industry regulations; Plan 3 needs to confirm whether the passenger has declared the photographic equipment at check-in. If not, it does not comply with the airline's regulations on "compensation for undeclared valuables as ordinary luggage." The probability reasoning model analyzes the execution risks of each plan. For example, Plan 2 relies on the authenticity of the maintenance invoice, which may result in passengers falsely reporting expenses. According to historical data, the probability of such risk is 15%. If the passenger fails to declare the item in Plan 3, it may lead to subsequent legal disputes, with a risk probability of 25%. The decision reasoning engine combines the compliance and risk assessment results, recommends Plan 2 first, and prompts the operator to verify the authenticity of the maintenance invoice and whether the passenger has declared the item in advance.

[0050] Execute step S43 to build the human-computer interaction layer. This layer designs a multi-view visualization panel to support the comparative analysis of processing solutions, regulatory clause tracing, and electronic signature confirmation functions. In the case of flight CA1203, after the operator logs in to the system, he / she can view the basic information of the case (passenger name, flight number, abnormality type, etc.) on the left side of the visualization panel. The middle area displays a comparison table of candidate processing solutions, with the solution number, compensation amount, processing time limit, and rule / case listed horizontally, and key dimensions (such as liability determination, compensation standard, and risk level) listed vertically. For example, Solution 1 compensates 500 yuan with a time limit of 3 working days, based on Article 8 of the Air China Baggage Transportation Rules; Solution 2 compensates 2,000 yuan with a time limit of 5 working days, based on case CA20230405 (damaged photographic equipment is compensated according to the repair cost); Solution 3 compensates 5,000 yuan with a time limit of 7 working days, based on Article 37 of the Civil Aviation Transport Regulations. The operator clicks the "Based on Rules / Case" link of any solution to jump to the regulatory clause details page or case archive page to view the specific content and effective time. A risk assessment chart is provided on the right side of the panel, which displays the scores of each plan in terms of compliance, execution cost, passenger satisfaction and other dimensions in the form of a radar chart.

[0051] In the process confirmation stage of the solution, the operator pushes the detailed description of Solution 2 to the passenger through the visual panel, and the passenger online confirms the processing solution through the electronic signature function. The system automatically generates an electronic confirmation letter, including the solution content, the rights and obligations of both parties, the compensation payment method, etc. The electronic signature is stored using blockchain technology to ensure non-tampering. If the passenger has any objections to the solution, they can initiate a negotiation process through the panel, and the operator can readjust the retrieval parameters or activate the manual review mechanism. For example, if the passenger states that the photographic equipment is a limited edition and the repair cost exceeds 2,000 yuan, the operator can add retrieval keywords such as "limited edition items" and "repair cost exceeding expectations" to the vector index library, retrieve similar cases again, and obtain solutions for handling special items.

[0052] During the system integration process, the hybrid retrieval engine module, the decision-making inference engine module, and the human-machine interaction layer achieve data interaction through API interfaces. For example, after the human-machine interaction layer receives the case description input by the user, it calls the retrieval interface of the hybrid retrieval engine through the API to obtain a set of candidate solutions; then it calls the evaluation interface of the decision-making inference engine through the API to obtain the compliance and risk scores of the solutions; finally, it returns the processing results to the human-machine interaction layer for display through the API. Each module is deployed on a cloud server and adopts a microservices architecture, supporting dynamic expansion and load balancing to ensure the stability of the system in high-concurrency scenarios.

[0053] The system also has a log management function, which records the entire process of each case's processing flow, including retrieval parameters, candidate solution generation time, decision-making inference process, operator approval records, passenger confirmation time, etc. The log data is stored in an independent log server and supports querying by dimensions such as case number, processing time, and operator, facilitating subsequent auditing and process optimization. For example, through log analysis, the management can find that the average processing time for a certain type of abnormal luggage case is 48 hours, which is higher than the industry standard of 24 hours, and then investigate the bottleneck links in the retrieval parameter settings or decision-making process, and adjust the system configuration to improve efficiency.

[0054] Through the above steps, the abnormal luggage disposal system realizes the full-process automated processing from case information entry, hybrid retrieval, decision-making inference to solution confirmation. Each module has a clear division of labor and close cooperation, ensuring the efficiency, accuracy, and traceability of the processing process. The operator does not need to manually consult a large number of policy documents and historical cases, and can quickly obtain comprehensive processing suggestions based on rules and cases through the system. At the same time, with the help of visualization tools, the operator can clearly master the key nodes and basis of case processing, greatly improving the standardization and intelligence level of abnormal luggage disposal.

[0055] It should be noted that in this text, relational terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the term "comprising", "including" or any other variant thereof is intended to cover non-exclusive inclusion, so that a process, method, article or device comprising a series of elements not only includes those elements, but also includes other elements not expressly listed, or also includes elements inherent to such process, method, article or device.

[0056] Although the embodiments of the present invention have been shown and described, those of ordinary skill in the art can understand that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of the present invention. The scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. An abnormal luggage handling method based on vector matching, characterized in that The following steps are involved: Step S1: Constructing a hybrid retrieval knowledge base: integrating multi-source heterogeneous data of abnormal baggage cases to generate a relational knowledge base and a vector knowledge base; indexing and associating the relational knowledge base with the vector knowledge base to establish a hybrid retrieval knowledge base; Step S2: Establish a hybrid search rule engine: perform keyword analysis and semantic vectorization on the input abnormal baggage case description to generate a hybrid search query instruction; perform parallel search based on the hybrid search knowledge base to generate a set of candidate processing solutions; Step S3: Develop a domain-enhanced decision system based on a large model: enhance the candidate processing solution set with domain knowledge and generate structured processing suggestions; Output processing flow specifications and compensation basis through a visual interactive interface; Step S4: Build an abnormal baggage handling system: integrate the hybrid retrieval rule engine, domain-enhanced decision-making system and visual interactive interface to realize the full-process automated processing of abnormal baggage cases from retrieval to decision-making; Step S2 includes the following steps: Step S21: Parse the input case description, extract the airline code, exception type, and item feature fields, and generate a keyword search expression; Step S22: semantically encode the unstructured text to generate a query vector; calculate the cosine similarity between the query vector and the vectors in the vector knowledge base, and select the top N similar cases by ranking by score; Step S23: Merge the keyword search results and the vector search results, perform conflict resolution processing, and generate a candidate processing solution set including processing procedures, compensation standards, and historical cases.

2. The abnormal baggage handling method based on vector matching according to claim 1, wherein Step S1 includes the following steps: Step S11: Build a relational knowledge base: collect abnormal baggage handling policy documents of different airlines, perform document structure processing through localized parsing tools, and generate a triple data set of airline-abnormal type-handling rule; Step S12: constructing a vector knowledge base: collecting historical abnormal baggage case data, cleaning missing values ​​and standardizing fields, generating text vectors and multimodal joint vectors through a pre-trained language model, and storing them in a vector index library; Step S13: Establish an index association mechanism: set the airline code primary key for the relational knowledge base, configure a dynamic shard index for the vector knowledge base, and establish a cross-modal query channel for the hybrid retrieval knowledge base through primary key mapping.

3. The abnormal luggage processing method based on vector matching according to claim 1, wherein Step S22 includes the following steps: Step S221: Perform word segmentation and entity recognition on the case description text to extract airline identifier, abnormal item category, and damage level features; Step S222: using a pre-trained language model to generate a text semantic vector, performing feature dimensionality reduction through a multi-layer perceptron, and generating a query vector of fixed dimension; Step S223: perform shard parallel search in the vector index library, use weighted cosine similarity algorithm to calculate similarity score, merge the shard results and take the top N similar cases.

4. The abnormal baggage handling method based on vector matching according to claim 1, characterized in that, Step S3 includes the following steps: Step S31: Execute domain knowledge enhancement: extract airline-specific processing rules from the relational knowledge base, retrieve similar case processing records from the vector knowledge base, and form a multi-dimensional evidence chain; Step S32: Construct a RAG inference framework: Convert the retrieval results into structured prompt words and input them into the large language model to generate processing suggestions, including liability determination logic and compensation amount calculation basis; Step S33: Design a visual output template: Convert the processing suggestions into a structured flowchart, mark the key decision nodes and regulatory reference sources, and generate an editable disposal plan document.

5. The abnormal luggage processing method based on vector matching according to claim 4, characterized in that Step S31 includes the following steps: Step S311: Establish a rule weight matrix: Assign weight coefficients to evidence from different sources according to the airline policy priority and case confidence. Step S312: Perform evidence fusion calculation: Use a Bayesian network model to integrate structured rules and unstructured case data to generate a comprehensive processing suggestion confidence score. Step S313: Construct a decision tree model: Convert the fusion result into a decision tree structure including abnormal type judgment, compensation standard application, and processing flow path.

6. The abnormal luggage processing method based on vector matching according to claim 1, wherein Step S4 includes the following steps: Step S41: Deploy a hybrid retrieval engine module: Configure a multi-threaded concurrent controller to achieve synchronous query scheduling of a relational database and a vector index library. Step S42: Develop a decision inference engine: Integrate a rule engine and a probability inference model to support compliance verification and risk assessment of the processing plan. Step S43: Construct a human-computer interaction layer: Design a multi-view visualization panel to support functions such as comparison and analysis of processing plans, traceability of regulatory clauses, and electronic signature confirmation.

7. The abnormal luggage handling method based on vector matching according to claim 2, wherein In step S12, an adversarial training mechanism is introduced in the construction of the vector knowledge base, and adversarial luggage damage image data is generated through the GAN network.

8. The abnormal luggage processing method based on vector matching according to claim 1, wherein The domain-enhanced decision-making system in step S3 includes a dynamic knowledge distillation module: Establish a three-layer architecture including a basic rule layer, a case reasoning layer, and a meta-learning layer, and gradually increase the training weight of complex cases through a curriculum learning strategy.

9. An abnormal luggage handling system based on vector matching, characterized in that, Including: A hybrid retrieval knowledge base construction module for integrating multi-source heterogeneous data of abnormal luggage case data to generate a relational knowledge base and a vector knowledge base, and associating the two through indexing to form a hybrid retrieval knowledge base; A hybrid retrieval rule engine module for keyword parsing and semantic vectorization of the input abnormal luggage case description to generate a hybrid retrieval query instruction, and performing parallel retrieval based on the hybrid retrieval knowledge base to output a candidate processing plan set; specifically including: Parse the input case description, extract the airline code, abnormal type, and item feature fields, and generate a keyword retrieval expression; Semantically encode the unstructured text to generate a query vector; calculate the cosine similarity between the query vector and the vectors in the vector knowledge base, and screen the top N similar cases according to the score ranking; Merge the keyword retrieval results and the vector retrieval results, perform conflict resolution processing, and generate a candidate processing plan set including processing flow, compensation standard, and historical cases; A domain-enhanced decision-making module integrating a large language model for domain knowledge enhancement of the candidate processing plan set to generate structured processing suggestions, and outputting processing flow specifications and compensation basis through a visual interaction interface; An integration module for logically integrating the hybrid retrieval knowledge base construction module, the hybrid retrieval rule engine module, and the domain enhancement decision module to achieve full-process automated processing of abnormal luggage cases from retrieval to decision-making.

Images