Airport luggage abnormity claim settlement process monitoring method based on artificial intelligence

Abstract

The invention discloses an airport luggage abnormity claim settlement process monitoring method based on artificial intelligence. The method comprises the following steps: firstly, constructing and dynamically maintaining a multi-dimensional luggage transportation space-time knowledge graph STKG-Bag; secondly, a multi-modal attention fusion deep network M2ARD-Net is utilized to carry out intelligent analysis and breakpoint positioning on a responsibility chain of a luggage abnormal event; thirdly, when the luggage track is fuzzy, intelligently reconstructing a luggage path through dynamic track reconstruction based on deep reinforcement learning and a multi-party responsibility quantification process DRL-CRQ, and adaptively optimizing a multi-party compensation proportion by using a graph neural network GNN-Pay; and finally, supporting a decision through an interactive visual interface, and realizing closed-loop feedback to continuously optimize each AI model and the knowledge graph. According to the method, the dynamic knowledge graph, multi-modal deep learning and reinforcement learning technologies are fused, automation, precision, high efficiency and fairness of luggage abnormal claim settlement are realized, and the processing efficiency and the passenger experience are remarkably improved.

Description

Technical Field

[0001] The present invention belongs to the field of airport operation management, and in particular relates to an airport baggage abnormal claims process monitoring method based on artificial intelligence. Background Art

[0002] Airport baggage transportation is a core link in the civil aviation industry, which directly affects the passenger experience and the operating efficiency of airlines. The accuracy and efficiency of its abnormal claims process are crucial to improving service quality. With the complexity of the global air transport network, abnormal baggage incidents occur frequently, involving multiple carriers and complex handover links. It has become extremely difficult to define the responsibility for abnormal baggage (such as loss, damage, and delay). The traditional claims process is highly dependent on manual intervention, which is not only inefficient and costly, but also prone to reduced passenger satisfaction and disputes between airlines due to factors such as information asymmetry, broken evidence chain, and subjective judgment differences.

[0003] Although existing baggage tracking systems (such as systems based on BSM messages) can provide some baggage flow data, these systems generally have the following limitations: First, data standards are not unified among the systems of airlines, airports, and ground agents, making information sharing difficult and difficult to form a complete, end-to-end view of baggage transportation; second, scanning data from key nodes may be delayed in uploading or completely missing, especially in manual operation links or areas not covered by the system, resulting in incomplete baggage tracks; third, when anomalies occur, it is impossible to automatically identify potential breakpoints in the chain of responsibility, and it is impossible to perform deep mining and intelligent reasoning when data is missing.

[0004] The development of deep learning and knowledge graph technology has provided new opportunities for solving such complex problems. However, how to build a comprehensive solution that can adapt to the dynamic characteristics of air baggage transportation, effectively handle data uncertainty, and provide explainable intelligent decision-making is still a technical bottleneck that needs to be broken through in this field. Summary of the invention

[0005] The present invention aims to overcome the technical problems existing in the existing airport baggage abnormality claims process, such as fuzzy liability determination, lengthy processing cycle, strong subjectivity in decision-making, difficulty in cross-subject collaboration, and insufficient intelligence level. A comprehensive monitoring and intelligent decision-making method is provided that integrates dynamic knowledge graph construction, multimodal deep learning-driven liability boundary identification, fuzzy trajectory intelligent reasoning based on deep reinforcement learning, and adaptive compensation plan generation, so as to realize the automation, precision, efficiency and fairness of baggage abnormality event processing.

[0006] The present invention discloses an artificial intelligence-based method for monitoring abnormal baggage claims process at an airport, comprising the following steps: Collect and fuse in real time the baggage transportation data from multi-source heterogeneous systems, and construct and dynamically maintain the spatio-temporal knowledge graph STKG-Bag of baggage transportation through knowledge extraction, entity alignment and relationship inference. The STKG-Bag contains entities and the spatio-temporal association relationships between entities. The entities include baggage, passengers, flights, carriers, operators, geographical locations, time points, operation events and abnormal events; When a baggage abnormal event is triggered, extract the spatio-temporal sub-graph and operation event sequence related to the abnormal baggage from the STKG-Bag, and input the spatio-temporal sub-graph and operation event sequence into the pre-trained multi-modal attention fusion deep network M 2 ARD-Net. The M 2 ARD-Net encodes the graph structure context, temporal event sequence, and prior knowledge of operation specifications and abnormal patterns, and uses multi-modal feature fusion and hierarchical attention mechanism to obtain the analysis result of the breakpoint of the responsibility chain; If the analysis result of the breakpoint of the responsibility chain shows that the actual trajectory of the baggage is unclear in a specific area, start the dynamic trajectory reconstruction and multi-party responsibility quantification DRL-CRQ process based on deep reinforcement learning to obtain the reconstructed path and compensation plan; Interactively display the baggage trajectory, the analysis result of the breakpoint of the responsibility chain, the reconstructed path and the compensation plan in the STKG-Bag through an integrated multi-dimensional visualization interface. After receiving the final claim settlement plan input by the user, use the final claim settlement plan to perform continuous incremental learning and iterative optimization on the STKG-Bag, M 2 ARD-Net model, DRL-CRQ agent and GNN-Pay model.

[0007] Preferably, knowledge extraction of baggage transportation data includes extracting entities and events from unstructured text by using a pre-trained language model based on Transformer combined with conditional random field or pointer network. Relationship inference includes using graph neural network for link prediction.

[0008] Preferably, the spatio-temporal context encoding module STCE-Module in the M 2 ARD-Net includes: a graph structure context encoder GSC of a graph attention network GAT for encoding the topological structure of the baggage transportation sub-graph and the dependencies between nodes; a temporal event sequence encoder TESE composed of a bidirectional gated recurrent unit Bi-GRU with self-attention mechanism for capturing the long-term dependencies and time dynamics of the operation event sequence; and a prior knowledge encoder PKE-Module for encoding operation specifications and historical abnormal patterns as prior knowledge and aligning them with the current event features.

[0009] Preferably, the M 2 ulti-modal Feature Fusion and Hierarchical Attention Module M of ARD-Net 2 The FA-Module uses a gated fusion unit to perform early fusion on the multi-modal features from the STCE-Module, then realizes in-depth interactive learning between features through a cross-modal Transformer encoder, and applies a hierarchical fine-grained attention mechanism to locate anomalies from the macroscopic section to the microscopic event.

[0010] Preferably, the M 2 Responsibility Attribution and Breakpoint Confidence Evaluation Module RAD-Module of ARD-Net uses a multi-task learning head to parallelly predict the probability of the breakpoint position, the anomaly cause code, and the preliminary tendency score of the responsible entity, and highlights the input features that contribute greatly to the decision-making by combining interpretability methods.

[0011] Preferably, start the dynamic trajectory reconstruction and multi-party liability quantification DRL-CRQ process based on deep reinforcement learning to obtain the reconstructed path and the compensation plan, specifically including: Construct a dynamic environment that simulates the baggage transfer based on STKG-Bag, operation rules, and breakpoint information; Through a deep reinforcement learning agent, use the Actor-Critic framework and combine Monte Carlo tree search to assist exploration to perform policy learning in the environment, and reconstruct N most likely baggage movement paths and their probabilities according to the refined designed reward function; For each high-probability reconstructed path, construct an interaction graph of the responsible parties, and use the graph neural network compensation ratio allocation model GNN-Pay to quantify the compensation responsibilities of each relevant party, and output an adaptive multi-party compensation ratio allocation plan; Among them, the reward function of the deep reinforcement learning agent comprehensively considers path coherence, time consistency, interpretability of anomaly recognition, satisfaction with operation rules and physical constraints, conformity with historical data patterns, and path complexity.

[0012] Preferably, the input of the graph neural network compensation ratio allocation model GNN-Pay is the interaction graph of the responsible parties constructed based on the high-probability reconstructed path, where the node features include the preliminary liability tendency score of the relevant party, the participation degree on the reconstructed path, the historical operation error rate, and the compliance with norms. The model outputs the compensation percentages of each relevant party and is trained by supervised learning of historical cases and industry rule constraints.

[0013] Preferably, the continuous incremental learning and iterative optimization include adopting an active learning strategy to select high-value samples for manual annotation, and using the newly confirmed cases to evolve and update the knowledge graph, for M 2The ARD-Net, DRL-CRQ agent, and GNN-Pay model perform periodic or event-triggered incremental training and parameter fine-tuning.

[0014] The present invention also discloses an artificial intelligence-based monitoring system for the abnormal claim settlement process of airport luggage, including a memory and a processor. The memory stores a computer program, and when the processor executes the computer program, the above-mentioned artificial intelligence-based monitoring method for the abnormal claim settlement process of airport luggage is implemented.

[0015] The present invention also discloses a computer-readable storage medium, on which a computer program is stored. The computer program, when executed by a processor, implements the above-mentioned artificial intelligence-based monitoring method for the abnormal claim settlement process of airport luggage.

[0016] Compared with the prior art, the present invention constructs a unified STKG-Bag by deeply integrating a dynamic knowledge graph, which helps to break data islands and promote trust and efficient cooperation among airlines, airports, and ground services. When there is data loss or conflict, the DRL-CRQ engine can intelligently deduce the most likely luggage path based on constraints and historical experience, providing key evidence for liability attribution. By constructing M 2 Through multi-modal information fusion and hierarchical attention, ARD-Net can identify the real breakpoints of the liability chain from complex associations and subtle signals, improving the accuracy of liability determination. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 is a detailed architecture diagram of the multi-modal attention fusion deep network M 2 ARD-Net for intelligent analysis and breakpoint location of the liability chain in the embodiments of the present invention.

[0018] Figure 2 is the overall flowchart of the artificial intelligence-based monitoring and intelligent decision-making method for the abnormal claim settlement process of airport luggage in the embodiments of the present invention.

[0019] Figure 3 is a visualization effect diagram of the STKG-Bag subgraph, highlighted liability chain breakpoints, multiple fuzzy trajectories reconstructed by DRL, and GNN-Pay claim settlement plan suggestions in a complex multi-carrier transfer luggage abnormality case in the embodiments of the present invention.

[0020] Figure 4 is a simulation comparison bar chart of the present invention's method with traditional manual methods and methods based on simple rule systems in terms of key performance indicators (liability determination accuracy rate, average processing time, claim settlement plan adoption rate, passenger satisfaction index). DETAILED DESCRIPTION OF THE EMBODIMENTS

[0021] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0022] This embodiment discloses an artificial intelligence-based method for monitoring the abnormal claim settlement process of airport luggage (for its overall process, see the attached Figure 2 figure), which may specifically include: Step S1: Construct and dynamically maintain a multi-dimensional luggage transportation spatio-temporal knowledge graph (STKG-Bag) S1.1 Multi-source heterogeneous data real-time acquisition and fusion engine: Design an extensible data access and fusion framework to obtain multi-modal data from the Global Distribution System (GDS), airline departure control system (DCS), baggage source message (BSM), baggage handling system (BHS) logs, baggage reconfirmation system (BRS) scan records, flight information display system (FIDS), World Tracer, airport operations database (AODB), ground agent operation platform, and even authorized Internet of Things sensors (such as low-power wide-area network trackers on baggage tags) in real-time or near real-time. The data includes: baggage tag information (IATA standard message), passenger itinerary (PNR), flight schedule and status, baggage scan timestamp and location (accurate to sorting ports, apron positions, cargo holds), operator ID, electronic images of handover documents, abnormal report texts, etc.

[0023] S1.2 STKG-Bag and Schema Design: Define a refined spatio-temporal knowledge graph ontology for baggage transportation, including core entity types: Baggage, Passenger, FlightInstance, Segment, Carrier, Airport, Operator, GroundHandler, HandlingDevice, GeoLocation, Timestamp, HandlingEvent, ExceptionEvent, etc. Define rich attributes for each type of entity (e.g., Baggage.tag_number, FlightInstance.actual_departure_and_arrival_time, HandlingEvent.handling_type, ExceptionEvent.report_reason_code). Define complex semantic relationships: such as Baggage.transported_on.Segment, Segment.operated_by.Carrier, HandlingEvent.occurred_at_loc.GeoLocation, HandlingEvent.performed_by.Operator, ExceptionEvent.related_to_baggage.Baggage, as well as temporal order relationships, causal relationships, co-occurrence relationships, etc.

[0024] S1.3 Knowledge Extraction and Graph Instantiation: S1.3.1 Structured Data Mapping: Directly map structured data from sources such as databases and API interfaces to the entities and relationships of STKG-Bag.

[0025] S1.3.2 Semi-Structured / Unstructured Data Extraction: Use pre-trained language models based on Transformer (such as BERT, ERNIE) combined with Conditional Random Fields (CRF) or Pointer Networks to extract key entities (such as damaged parts with specific descriptions, handover details mentioned by passengers) and events from baggage exception reports, operation remarks, and handover document pictures (after OCR).

[0026] S1.3.3 Entity Alignment and Disambiguation: Handle the alignment of the same entity from different systems (such as different representations of the same flight in different systems), and utilize entity alignment algorithms based on graph embedding or disambiguation strategies based on rules. First, for different types of entities, define their core recognition attribute sets. For example, for "flight", use "flight number + scheduled departure date + three-letter code of departure location + three-letter code of destination" as features; for "passenger", use "(hashed) document number + surname" or "frequent flyer number" as features. Generate preliminary candidate entity pairs using exact matching or edit distance (such as the Jaro-Winkler algorithm, setting a similarity threshold, for example, above 0.9).

[0027] Secondly, utilize the relationships between entities already constructed in STKG-Bag for verification and weight adjustment. For example, if two preliminarily matched "passengers" are respectively associated with "baggages" having the same "baggage tag number", the confidence level of the alignment of these two "passengers" is significantly improved. Conversely, if there are conflicts in the associated flight information, the confidence level is reduced.

[0028] Thirdly, apply specific rules in the field of baggage transportation for forced alignment or exclusion. For example, according to IATA Resolution 753, baggage records with the same 10-digit baggage tag serial number (License Plate Number) should be highly prioritized for alignment as the same piece of baggage even if other information varies slightly.

[0029] Finally, for alignment results with low confidence or conflicting alignments, an artificial review mechanism can be introduced, and the review results can be fed back to the alignment model for learning to achieve iterative optimization.

[0030] Specifically, the entity alignment process preferably includes the following steps: (1) feature matching based on key attributes; (2) context relationship verification and weighting; (3) domain-specific rule prioritization; (4) iteration and artificial review.

[0031] S1.3.4 Relationship Inference and Link Prediction: Utilize graph neural networks (GNNs) such as GraphSAGE or RotatE and other models to train on the constructed graph fragments to infer potential missing relationships or predict future possible links (such as predicting which unit load device (ULD) the baggage is most likely to be loaded into).

[0032] S1.4 Graph Dynamic Update and Version Control: Ensure that STKG-Bag can nearly real-time reflect the latest status of the baggage and changes in the transportation environment. Introduce an event sourcing mechanism and the version control function of the graph database to track historical states and conduct retrospective analysis.

[0033] Step S2: Intelligent Dissection and Breakpoint Location of the Chain of Responsibility Based on the Multi-modal Attention Fusion Deep Network (M 2 ARD-Net) When a baggage anomaly event (such as a Baggage Irregularity Report, PIR) is triggered, extract all historical operation event sequences related to the abnormal baggage from STKG-Bag, including the involved carriers, transfer points, timestamps, operators, etc., to form a spatio-temporal subgraph. Construct a multi-modal attention fusion deep network (M 2 ARD-Net, Multi-modal Attentive Responsibility Dissection Network), whose detailed network architecture is shown in the appendix Figure 1 to accurately locate the breakpoint of the chain of responsibility and preliminarily evaluate the abnormal contribution degree of each link.

[0034] S2.1 Spatio-temporal Context Encoding Module (STCE-Module): S2.1.1 Graph Structure Context Encoder (GSCE): Use a multi-layer graph attention network (GAT) to process the extracted baggage transportation subgraph. GAT can learn the importance of nodes (such as carriers, transfer points, operators) in the graph structure and capture the complex interdependencies between them. The initial features of the nodes can be composed of their embedding vectors or attributes in STKG-Bag. Output the context-aware representation of each node.

[0035] S2.1.2 Temporal Event Sequence Encoder (TESE): Encode the time-ordered baggage operation event sequence (each event contains features such as time, location, operation type, operator ID, etc.). Use a bidirectional gated recurrent unit with self-attention mechanism (Bi-GRU with Self-Attention) network. Bi-GRU captures the long-term dependencies and temporal dynamics between events, and the self-attention mechanism helps the model focus on the few events that are most critical to the occurrence of the anomaly in the sequence.

[0036] S2.1.3 Prior Knowledge Encoder for Operation Specifications and Anomaly Patterns (PKE-Module): a. Embedding of the operation specification library: Structurally represent the key rules in the airline's standard operating procedures (SOPs) and the IATA Baggage Handling Manual (BHM) (such as rule graphs or vector embeddings) to form an operation specification knowledge base.

[0037] b. Historical Anomaly Pattern Library: Extract typical anomaly patterns (such as "when the transfer time of a specific transfer station for the A-B route is less than 30 minutes, the misplacement rate of luggage increases significantly") from the historical resolved luggage anomaly cases through clustering or pattern mining algorithms (such as Apriori combined with sequence pattern mining).

[0038] c. Prior Knowledge Fusion: Use a small attention network to match and align the features of the current event sequence with the relevant entries in the operation specification library and the historical anomaly pattern library, generating feature vectors of "specification deviation degree" or "historical pattern matching degree". For example, the rule "when the transfer time is less than the minimum connecting time (MCT) specified by the airport, the fast transfer procedure should be initiated" can be represented as a structured object containing conditions (transfer time < MCT), context (specific airport / carrier), and expected actions (initiate fast transfer), and its text description or structured features are encoded as dense vectors through a pre-trained language model (such as BERT). For example, an excavated anomaly pattern may be: "Event sequence: [Flight A arrives at transfer station X] -> [Luggage is unloaded and scanned at station X] -> [No subsequent loading scan for a long time (such as more than 1.5 times the average transfer time of station X)] -> [The passenger reports lost luggage at destination Y]". These extracted patterns are also transformed into feature vectors or graph patterns for similarity matching with the current event sequence.

[0039] S2.2 Multi-modal Feature Fusion and Hierarchical Attention Module (M 2 FA-Module): S2.2.1 Early Fusion and Gating Mechanism: Initially fuse the graph-structured context representation output by GSCE, the temporal event representation output by TESE, and the prior knowledge representation output by PKE through a learnable gating mechanism (Gated Fusion Unit), which can dynamically adjust the weights of different modal information.

[0040] This gating unit dynamically weights the relative importance of different information in the luggage claim scenario. For example, in a scenario where the luggage track record is complete but the liability determination is ambiguous, the compliance analysis of operation specifications provided by the PKE module may obtain a higher weight; while in a scenario where key data in the luggage track is missing, the historical behavior patterns of relevant entities (such as carriers, operators) provided by GSCE and the event sequence context provided by TESE may be more important.

[0041] S2.2.2 Cross-modal Interaction Attention Layer: Based on the fused features, multiple layers of cross-modal Transformer encoders are applied. Through self-attention and feed-forward network layers, this encoder enables in-depth interaction of features from different sources (such as temporal dimension features and spatial / structural dimension features), learns complex non-linear associations between them, and thus identifies subtle abnormal signals hidden in multi-modal data.

[0042] S2.2.3 Hierarchical Fine-grained Attention: Design a hierarchical attention mechanism. First, allocate attention at the macro level (such as flight segments, main transfer points) to identify the most suspicious responsible sections; then, within the identified suspicious sections, further refine the attention to specific individual operation events or time points to achieve abnormal localization from coarse to fine.

[0043] S2.3 Responsibility Attribution and Breakpoint Confidence Evaluation Module (RAD-Module): S2.3.1 Multi-task Breakpoint Prediction Head: The final fused feature representation is fed into the multi-task learning head, which concurrently performs the following tasks: a. Breakpoint Location Classification: For each operation event or handover link, predict the probability of whether it is a breakpoint in the responsibility chain (binary classification or multi-class classification if there are types of breakpoints).

[0044] b. Abnormal Cause Code Prediction: Predict the most likely cause code leading to the breakpoint (such as IATA standard cause codes: baggage loading error, baggage offloading error, documentation error, transfer mishandling, etc.). The abnormal cause code system adopted in the present invention is based on the IATA standard baggage accident cause codes (such as the codes in IATA AHM Chapter 7, for example, 52 - Transfer Mishandled, 73 - Failure to Load at Transfer), and is appropriately refined and extended according to actual operation experience and data statistics to form a hierarchical classification system. For example, "73 - Failure to Load at Transfer" is further divided into sub - classes such as "73.1 - ULD not loaded as planned", "73.2 - Baggage sorting error and not delivered to the designated ULD", "73.3 - Baggage has been delivered to the ULD but not loaded due to lack of space / time out", etc., in order to more accurately locate the problem. Give a preliminary abnormal relevance or liability tendency score to each responsible entity (such as carrier, ground service agent, airport operation department) involved in this breakpoint link. The basis of this score mainly comes from the patterns learned by the model from the input features. For example, whether the last operation of this entity before the breakpoint conforms to the operation specifications encoded by the PKE module, the historical failure rate of this entity recorded in STKG - Bag for similar scenarios, the actual control situation of this entity over the key control points (such as baggage handover) in the current event chain, etc. The score result is usually normalized to the interval [0, 1], indicating the association strength or potential liability size of this entity with the current abnormal event.

[0045] c. Preliminary Liability Tendency Score for Responsible Entities: Give a preliminary abnormal relevance score to each responsible entity (carrier, ground service) involved in this link.

[0046] S2.3.2 Interpretability Analysis: Interpretability AI methods such as Integrated Gradients or LIME are used to highlight the parts of the input features that contribute the most to breakpoint determination and cause prediction (such as a missing scan record, an abnormal time interval, or an irregular operator signature), providing a basis for manual review. The output is a structured breakpoint report that includes the breakpoint location, confidence level, predicted cause, preliminary responsible party tendency, and supporting evidence fragments. For example, when the model mainly attributes the responsibility of "baggage not loaded onto the subsequent flight" to the transfer ground service, the interpretability analysis may highlight two features: "in the input event sequence, the operation time of this ground service in the transfer baggage handover link far exceeds the average level" and "the prior knowledge of 'the human resources of this ground service are tense during this period' matched by the PKE module", thus assisting manual reviewers in understanding the decision-making logic of the model.

[0047] Step S3: Intelligent Reconstruction of Fuzzy Trajectory and Multi-party Liability Quantification Based on Deep Reinforcement Learning and Constraint Programming (DRL-CRQ) When significant breakpoints in the responsibility chain are identified in Step S2 (such as missing key scan records or severely abnormal handover times), resulting in the actual trajectory of the baggage being unclear in a certain area, this step is initiated.

[0048] S3.1 Dynamic Environment Construction and State Space Definition: S3.1.1 Environment: Based on the confirmed baggage trajectory segments in STKG-Bag, airport facility layout (from AODB or GIS), flight schedules, carrier / ground service operation capacity constraints (such as unit time processing capacity, standard transfer time window), and the breakpoint information output by S2 (as "signals" or "perturbations" in the environment), a simulated dynamic environment for baggage transfer is constructed.

[0049] S3.1.2 State Space (S): The state of the DRL agent includes: the currently hypothesized baggage location (discrete or continuous), the current time, a summary of the known historical trajectory of the baggage, the breakpoint information and confidence level given by S2, the congestion index of the surrounding environment (if available), and the satisfaction of the constraints of the current path. The airport facility layout (represented by a topological graph stored in a graph database or JSON format, with nodes being key processing units such as check-in islands, security check channels, baggage carousels, sorting slots, ULD building areas, and apron positions, and edges representing passable paths and their standard passing times, accurate to specific functional areas or key equipment), operation capacity constraints (for example, the maximum processing capacity of a certain sorting belt is N pieces per minute, and a certain ground service can handle M pieces of baggage per hour during the transfer peak period. These parameters serve as resource limitations and behavior constraints in the simulated environment).

[0050] S3.1.3 Action Space (A): The actions of the agent include: selecting the next possible physical location (such as a sorting slot, transfer storage area, ULD, cargo hold), selecting the next logical operation state (such as "waiting to be loaded", "in transit", "unloaded"), or "maintaining the current state and waiting" (simulating delays). The action space is dynamically generated based on the current location and airport operation rules. For example, if the currently simulated baggage is at the exit Z of the transfer sorting system and the SOP stipulates that exit Z only processes baggage for flights to region R1, then the available actions for the agent will only include directing the baggage to the container area or conveyor belt associated with flights in region R1, and actions leading to other regions will not be generated.

[0051] S3.2 Policy Network and Value Network Design (Based on the Actor-Critic Framework): Use the Actor-Critic algorithm (such as an improved version of A3C, Soft Actor-Critic (SAC), or Proximal Policy Optimization (PPO)).

[0052] S3.2.1 Actor Network (Policy Network): Given the current state S as input, it outputs the probability distribution of taking each possible action A in the current state (for a discrete action space) or a deterministic action (for a continuous action space). The policy network itself can be a deep neural network that includes convolutional layers (if the location is represented as an image), recurrent layers (to process temporal information in the state), and fully connected layers.

[0053] S3.2.2 Critic Network (Value Network): Given the current state S (and sometimes the action A) as input, it outputs a value assessment (Q-value or V-value) of the current state (or state-action pair), which is used to guide the learning of the Actor network.

[0054] S3.3 Fine-grained Design of the Reward Function: The reward function is the core that guides the agent to reconstruct the most "reasonable" baggage trajectory. Its design goal is to make the reconstructed trajectory be able to explain anomalies and conform to physical and operational logic. The reward function R is the sum of multiple weighted sub-items: R = w_coh * R_coherence + w_time * R_temporal + w_exp * R_explanation + w_con * R_constraint + w_hist * R_historical_match - w_comp * P_complexity.

[0055] Examples of the specific quantification methods for each sub-item are as follows: a. Path Coherence Reward R_coherence: If the reconstructed path segment can be smoothly connected to the known trajectory segments in the STKG-Bag both logically and physically at the start and end points (e.g., the spatial distance is less than the threshold ε), then R_coherence is a relatively large positive value (such as +20); otherwise, it is 0 or a negative value.

[0056] b. Time Consistency Reward R_temporal: For each simulated operation step in the path, if its elapsed time conforms to the standard operation time or flight schedule (such as MCT), a positive reward is given; if there is an unreasonably long stagnation or ultra-high-speed movement without reasonable explanation, a significant negative reward is given (e.g., a penalty of -10 is given if the time deviation exceeds 50% of the standard operation time).

[0057] c. Anomaly Explanation Reward R_explanation: If the reconstructed path can reasonably explain the breakpoints or anomalies identified in step S2 (e.g., the path shows that the luggage entered an area where misrouting often occurred in history before the breakpoint, or the path conforms to a certain known anomaly pattern encoded in the PKE module), then R_explanation is a relatively large positive value (such as +30).

[0058] d. Constraint Satisfaction Reward / Penalty R_constraint: If the path strictly adheres to airport operation rules (such as luggage not entering restricted areas like the customs supervision area), physical constraints (such as luggage cannot pass through walls), and luggage handling SOPs, then R_constraint is 0 or a small positive value; for each violation of a hard constraint, a relatively large negative penalty is superimposed (such as -100).

[0059] e. Historical Data Matching Reward R_historical_match: If the pattern of the reconstructed path (such as the sequence of key processing nodes, the combination of operator types involved) highly matches the pattern of the true historical luggage trajectories or high-confidence reconstructed trajectories recorded in the STKG-Bag with similar initial conditions (such as the same O&D, the same carrier combination, similar anomaly types) (e.g., measured using trajectory edit distance or sequence embedding cosine similarity, and the similarity is higher than the threshold 0.8), then R_historical_match is a positive value, and the higher the matching degree, the greater the reward.

[0060] f. Path Complexity Penalty P_complexity: To avoid generating overly long or unnecessarily complex circuitous paths, a slight negative penalty proportional to the path length (the number of operation steps) or the number of turns (the number of direction changes) is imposed on the reconstructed path (e.g., for each additional step or turn, a penalty of -0.1 is given).

[0061] Each weight coefficient w_coh, w_time, w_exp, w_con, w_hist, and w_comp is determined through multiple rounds of experiments and parameter tuning in the simulation environment.

[0062] S3.4 Monte Carlo Tree Search (MCTS) Aided Exploration and Planning: At certain decision points, especially when the action space is large or the environmental uncertainty is high, MCTS can be combined to assist the DRL agent in conducting deeper exploration and planning. Specifically, each iteration of MCTS includes four stages: selection, expansion, simulation, and backpropagation. In the selection and expansion stages, the action probability distribution output by the DRL's policy network (Actor network) can be used as the prior policy for MCTS tree search to guide the selection of nodes and the expansion of new nodes. In the simulation (Rollout) stage, the DRL's policy network can be used to quickly execute to the end, or the DRL's value network (Critic network) can be used to evaluate the future expected return of the current leaf node to truncate the simulation and reduce the computational cost. After MCTS completes a specified number of iterations, the visit counts or cumulative Q-values of each child node under the root node (corresponding to all optional actions in the current state) can be used to generate a better action selection strategy (e.g., select the action with the most visit counts), which can be directly used to guide the agent to make decisions in the current simulation step, or to generate high-quality (state - improved policy) sample pairs to further train and optimize the DRL's policy network.

[0063] S3.5 Multi - path Generation and Probability Evaluation: The DRL agent may explore and evaluate multiple candidate reconstruction paths. The system records N most likely paths and their corresponding cumulative rewards or probabilities given by the policy network.

[0064] S3.6 Liability Quantification and Compensation Ratio Intelligent Decision - making Based on the Reconstruction Path Set (GNN - Pay): S3.6.1 Constructing the Interaction Graph of Liability - related Parties: For each high - probability reconstruction path, identify all relevant liability entities (carriers, ground services, airport departments) on the path. Use these entities as nodes and their operational handovers or contractual relationships as edges to construct a small - scale interaction graph. Node features include: the preliminary liability tendency score assigned to the entity in S2, its participation in the current reconstruction path (such as operation duration, control of key nodes), its historical operation error rate (extracted from STKG - Bag), and whether it violates relevant operation specifications.

[0065] S3.6.2 Graph Neural Network Compensation Ratio Allocation Model (GNN-Pay): Design a graph neural network (such as the graph isomorphism network GIN or GAT with an attention mechanism), and input the above-mentioned interaction graph of responsible parties. The GNN can learn the responsibility allocation pattern under the complex interactions of multiple parties. The output layer of the model is a Softmax layer, which directly outputs the compensation percentages of each relevant responsible entity for the current abnormal event. When training this model, historical adjudicated compensation cases (including the liability ratios of each party) can be used as supervision signals, and its participation degree in the current high-probability reconstruction path (this participation degree is preferably quantified by one or a combination of the following methods: (i) the percentage of the number of operation steps responsible for execution by this entity in the total number of steps of the path; (ii) during the fuzzy time period covered by this reconstruction path, the simulated duration of the luggage under the actual control or management of this entity accounts for the percentage of the total fuzzy duration; (iii) for key turning points or decision points in the path, if the behavior of this entity directly causes this turn or decision, then its participation degree is increased by a preset weight), and industry rules such as the IATA Multilateral Interline Baggage Compensation Agreement (MIPA) are used as constraints or regularization terms. Specifically, in the loss function of the GNN model, in addition to the standard supervision loss (such as cross-entropy loss or mean squared error loss, used to fit the compensation ratios of historical cases), an additional regularization term is added. This regularization term is used to penalize the difference between the compensation ratios output by the model and the liability division ratios recommended by the MIPA rules in the applicable scenarios. For example, if MIPA stipulates that in a certain interline situation, the originating carrier bears X% of the liability and the transfer carrier bears Y% of the liability, then the regularization term can be expressed as: λ_mipa * (((Share_gnn_orig - X%)² + (Share_gnn_transfer - Y%)²), where Share_gnn_orig and Share_gnn_transfer are the compensation ratios of the originating and transfer carriers output by the GNN model respectively, and λ_mipa is a hyperparameter that controls the intensity of this regularization term. In this way, while learning the complex patterns of historical data, the GNN-Pay model can also ensure that its output results comply with the generally recognized fairness criteria in the industry to a certain extent.

[0066] S3.6.3 Multi-Path Result Fusion: If there are multiple high-probability reconstruction paths, the compensation ratios output by GNN-Pay under each path can be weighted and averaged (the weights are the probabilities or confidences of the paths) to obtain a final comprehensive compensation plan recommendation.

[0067] Step S4: Interactive Visualization, Decision Support, and Closed-Loop Feedback for Continuous Optimization S4.1 Multi-Dimensional Visualization and Intelligent Interpretation Dashboard: Incorporate the complete life cycle of the luggage in STKG-Bag (including the fuzzy part reconstructed in S3), M2 The breakpoints in the chain of responsibility identified by ARD-Net (highlighted with reasons), the multiple candidate trajectories generated by DRL-CRQ (indicating probabilities with different colors or transparencies), and the compensation plans recommended by GNN-Pay are visualized through an integrated web interface. The interface supports interactive exploration, such as clicking on a breakpoint to view a detailed analysis, simulating the impacts under different compensation plans, etc. A summary report generated in natural language is provided to explain the decision-making process and key bases.

[0068] S4.2 Human-Machine Collaborative Decision-Making and Audit Trail: Claims adjusters can make final decisions based on the intelligent analysis results of the system. The system should allow adjusters to intervene and modify key judgments (such as breakpoint confirmation, path selection, adjustment of compensation ratios), and record all manual operations and reasons to form a complete audit trail chain.

[0069] S4.3 Continuous Learning and Model Evolution Based on Feedback: S4.3.1 Active Learning and Annotation: For judgments with low system confidence or cases with a large number of manual corrections, trigger the active learning mechanism to request claims experts to provide high-quality annotations to expand the training dataset.

[0070] S4.3.2 Incremental Training and Fine-Tuning of the Model: Regularly (or after accumulating a certain amount of feedback data) use newly confirmed cases (including the results of manual corrections) to incrementally train or fine-tune the ARD-Net, DRL policy / value network, and GNN-Pay models to adapt to changes in the operating environment, newly emerging abnormal patterns, and updates to compensation rules. Adopt transfer learning techniques to effectively transfer the knowledge learned from old data to the new model. 2 ARD-Net, DRL policy / value network, and GNN-Pay models for incremental training or fine-tuning to adapt to changes in the operating environment, newly emerging abnormal patterns, and updates to compensation rules. Adopt transfer learning techniques to effectively transfer the knowledge learned from old data to the new model.

[0071] S4.3.3 Knowledge Graph Evolution: According to the claims settlement results and newly discovered patterns, dynamically update the ontology of STKG-Bag (such as adding new types of abnormal reasons) and instance data to ensure the freshness and accuracy of the knowledge graph.

[0072] To further clarify the specific operation process of the method described in the present invention and its application effects in actual scenarios, a typical complex multi-carrier transfer luggage anomaly case will be described in detail below.

[0073] This embodiment will demonstrate how the present invention gradually applies the core steps such as the construction of the spatio-temporal knowledge graph for luggage transportation, the intelligent analysis of the chain of responsibility, the intelligent reconstruction of fuzzy trajectories, the quantification of multi-party responsibilities, and the interactive decision support to achieve efficient and accurate handling of luggage anomaly events.

[0074] Assume a passenger's checked baggage (tag number LH789XYZ) with a planned itinerary: City S1 (Airport A, Carrier C1) -> City S2 (Airport B, where C1 and C2 code-share and transfer, and the ground agent GH_B is responsible for the actual transfer operation) -> City S3 (Airport C, Carrier C2). The passenger reports that the baggage has not arrived at Airport C.

[0075] Step S1: Construct and dynamically maintain a multi-dimensional baggage transportation spatio-temporal knowledge graph (STKG-Bag) S1.1 Multi-source heterogeneous data real-time acquisition and fusion engine: The system obtains the check-in record of LH789XYZ at Airport A from C1's DCS (time T0, operator OpA1, weight, destination C), and the BSM shows that the baggage has been loaded onto flight F1 (S1 - S2). After flight F1 arrives at Airport B, the system obtains the record of LH789XYZ being scanned in the BHS from the BRS at Airport B (time T1, location SortPier_X). The system obtains a record from the operation platform of GH_B, showing that the baggage LH789XYZ has been assigned to flight F2 (Carrier C2) flying to S3, and there is an electronic handover action (time T2, operator OpB1 signs for preparation to hand over to the loading bay of C2). No loading record of LH789XYZ on flight F2 is found in C2's DCS. WorldTracer has a PIR submitted by passenger P001 at Airport C, describing the characteristics of the baggage.

[0076] S1.2 STKG-Bag Ontology and Schema Design: Define a refined spatio-temporal knowledge graph ontology for baggage transportation, including core entity types: Baggage, Passenger, FlightInstance, Segment, Carrier, Airport, Operator, GroundHandler, HandlingDevice, GeoLocation, Timestamp, HandlingEvent, ExceptionEvent, etc. Define rich attributes for each type of entity (e.g., Baggage.tag_number, FlightInstance.actual_departure_and_arrival_time, HandlingEvent.handling_type, ExceptionEvent.reported_reason_code). Define complex semantic relationships: such as Baggage.transported_on.Segment, Segment.operated_by.Carrier, HandlingEvent.occurred_at_loc.GeoLocation, HandlingEvent.performed_by.Operator, ExceptionEvent.related_to_baggage.Baggage, as well as temporal order relationships, causal relationships, co-occurrence relationships, etc. Create entities according to the above predefined ontology: Baggage:LH789XYZ, Passenger:P001, FlightInstance:F1, FlightInstance:F2, Carrier:C1, Carrier:C2, GroundHandler:GH_B, Airport:A, Airport:B, Airport:C, HandlingEvent:CheckIn_A, HandlingEvent:Load_A_F1, HandlingEvent:Unload_B_F1, HandlingEvent:Sort_B, HandlingEvent:TransferPrep_B_F2, ExceptionEvent:PIR_C, etc. Attribute filling: Baggage:LH789XYZ.destination = C, HandlingEvent:CheckIn_A.timestamp = T0, ExceptionEvent:PIR_C.reported_reason = "Not_Received_At_Destination".

[0077] Relationship building: Baggage:LH789XYZ transported_on Segment(F1, A, B), Segment(F1, A, B) operated_by Carrier:C1, HandlingEvent:TransferPrep_B_F2 performed_by Operator:OpB1 (of GH_B), HandlingEvent:TransferPrep_B_F2 intended_forFlightInstance:F2.

[0078] S1.3 Knowledge extraction and graph instantiation: The text "The baggage is a blue hard shell with a yellow ribbon" in the PIR report is extracted by the NLP model as Baggage:LH789XYZ.color = "Blue", Baggage:LH789XYZ.shell_type = "Hard", Baggage:LH789XYZ.distinguishing_mark = "Yellow_Ribbon". If the operation platform record of GH_B is a scanned paper handover form, key information is extracted after OCR recognition. The system discovers that T_F2_dep - T2 < the minimum connecting transit time (MCT) based on the planned departure time T_F2_dep of flight F2 and the time T2 of TransferPrep_B_F2, which may be a potential risk point and is marked by the rule engine or the GNN link prediction module in the graph.

[0079] S1.4 Graph dynamic update: If C2 later finds LH789XYZ in a certain corner of airport B and supplements the scanned information, STKG-Bag will update the status and location of the baggage in real time.

[0080] Step S2: Based on M 2 Responsibility chain intelligent analysis and breakpoint location of ARD-Net Extract the spatio-temporal subgraph and event sequence related to LH789XYZ: "A_CheckIn (T0, C1)" -> "A_Load_F1 (C1)" -> "B_Unload_F1 (GH_B for C1)" -> "B_Sort (T1, GH_B)" -> "B_TransferPrep_F2 (T2, OpB1 of GH_B for C2)" -> [Missing C2 loading record] -> "C_PIR_Reported (P001)".

[0081] Among them, the extracted event sequences are specifically described as follows: Event “A_CheckIn (T0, C1)”: This event indicates that the luggage LH789XYZ was recorded as having completed check-in and consignment procedures at the origin airport A by the agent or system of carrier C1 at time point T0. In STKG-Bag, this event node is associated with the luggage LH789XYZ, passenger P001, flight F1 (subsequent connecting flight), carrier C1, geographical location airport A, and time point T0.

[0082] Event “A_Load_F1 (C1)”: This event indicates that the luggage LH789XYZ was recorded as having been loaded onto flight F1 operated by carrier C1 at the origin airport A. In STKG-Bag, this event node is associated with the luggage LH789XYZ, flight F1, carrier C1, and the corresponding operation time and location, and has a chronological order relationship with the event “A_CheckIn (T0, C1)”.

[0083] Event “B_Unload_F1 (GH_B for C1)”: This event indicates that after flight F1 arrives at the transit airport B, the luggage LH789XYZ was recorded as having been unloaded from flight F1, and this operation was performed by the ground agent GH_B on behalf of carrier C1. In STKG-Bag, this event node is associated with the luggage LH789XYZ, flight F1, ground agent GH_B (as the operation executor and indicating that it serves carrier C1), geographical location airport B, and the corresponding operation time and location.

[0084] Event “B_Sort (T1, GH_B)”: This event indicates that the luggage LH789XYZ was sorted at time point T1 at the transit airport B by the operator or automated system of the ground agent GH_B. In STKG-Bag, this event node is associated with the luggage LH789XYZ, ground agent GH_B, time point T1, and the specific sorting area within airport B.

[0085] Event “B_TransferPrep_F2 (T2, OpB1 of GH_B for C2)”: This event indicates that the luggage LH789XYZ was prepared for transfer at time point T2 at the transit airport B by the operator OpB1 of the ground agent GH_B, preparing to transfer it to the subsequent flight F2 operated by carrier C2. In STKG-Bag, this event node is associated with the luggage LH789XYZ, flight F2 (target transfer flight), operator OpB1 (affiliated with GH_B), carrier C2 (as the subsequent carrier), time point T2, and the corresponding operation location (such as the transit luggage collection and distribution area or the handover area).

[0086] Breakpoint "[Missing C2 Loading Record]": This is a potential breakpoint in the critical responsibility chain. In STKG-Bag, the system did not query or receive a valid record from carrier C2 regarding the loading of baggage LH789XYZ onto flight F2 (e.g., the corresponding BSM LDM message or DCS loading confirmation information is missing). This indicates that after the event "B_TransferPrep_F2 (T2, OpB1of GH_B for C2)", the transportation status of the baggage has become uncertain, which is a key focus for M²ARD-Net to conduct responsibility analysis and breakpoint location.

[0087] Event "C_PIR_Reported (P001)": This event indicates that passenger P001 submitted a baggage irregularity report (PIR) at the final destination airport C due to not receiving baggage LH789XYZ. In STKG-Bag, this event node (as an abnormal event entity) is associated with baggage LH789XYZ, passenger P001, the geographical location of airport C, and the report submission time, and serves as the trigger event for the entire abnormal claims process.

[0088] S2.1 STCE-Module: GSCE analyzes the handover relationship and historical cooperation records (if any in the graph) among C1, C2, and GH_B in this subgraph. TESE processes the event sequence and focuses on the lack of expected operations of C2 after T2. The PKE module matches the historical pattern of "the handover error rate between GH_B and C2 increases when the transit time at airport B is tight" (assuming it exists) and notes the deviation from the operation specification of T_F2_dep - T2 < MCT.

[0089] S2.2 M 2 FA-Module: The gating fusion unit weighs the graph structure information, temporal information, and prior knowledge. The cross-modal Transformer deeply interacts with these features. For example, it discovers that the "missing C2 record" in the time series is highly correlated with the "C1-GH_B-C2" handover chain in the graph structure and the "risk of insufficient transit time" in the prior knowledge. The hierarchical attention first focuses on the "transit link at airport B" and then refines to the specific operation of "GH_B handing over the baggage of flight F2 to C2".

[0090] S2.3 RAD-Module: Multi-task head output: - Breakpoint location: After HandlingEvent:TransferPrep_B_F2, in the loading link of FlightInstance:F2, confidence level 0.95.

[0091] - Abnormal cause code (predicted): IATA Code 52: Transfer Mishandled or IATA Code 73: Failure to Load at Transfer.

[0092] - Initial inclination of the responsible party: GH_B (the operator who executes the operation, may make mistakes due to time pressure) > C2 (failure to confirm receipt and loading). Interpretability analysis highlights: the absence of the BSM loading message of C2, the receipt record of operator OpB1 of GH_B at T2 has no corresponding receipt confirmation from C2, and the objective fact of insufficient transfer time.

[0093] Step S3: Intelligent reconstruction of fuzzy trajectories based on DRL-CRQ and quantification of multiple-party responsibilities Due to the lack of C2 loading records, the trajectory of the luggage is fuzzy after T2 and before the departure of flight F2.

[0094] S3.1 Dynamic environment construction: The environment simulates the operation area of GH_B at airport B, the container loading area / loading port of flight F2 of C2, and possible error flow areas. The state includes the current assumed position of the luggage, time, remaining time until the departure of F2, and the breakpoint signal identified in S2.

[0095] S3.2 Policy network and value network (SAC algorithm): The Actor network learns the policy of moving the luggage to the next position or operation state under different states. The Critic network evaluates the quality of this policy.

[0096] S3.3 Reward function: - If the path shows that the luggage is correctly delivered to the loading port of C2 before the departure of F2 and there is a simulated loading action, a high positive reward is given.

[0097] - If the path shows that the luggage is not delivered to C2 due to an operation error of GH_B (such as sending it to the wrong sorting port or forgetting it in the temporary storage area), and the time coincides with the cut-off receipt time of flight F2, an explanatory positive reward is given. - If the path violates certain operation regulations of airport B (such as entering a non-luggage area), a large negative penalty is given.

[0098] - If the path coincides with the historical pattern of "insufficient transfer time leading to chaotic operations of GH_B", a reward is given.

[0099] S3.4 MCTS assistance: When there are multiple possible flow directions in the operation area of GH_B, MCTS helps explore the optimal several moving directions.

[0100] S3.5 Multi-path generation: DRL generation example: - Path 1 (P = 0.60): After the luggage arrives at T2, it is sent by OpB1 to the ULD container area of Flight F2. However, due to the tight schedule or the busy receiving staff on the C2 side, the luggage is left at the edge of the container area and not loaded into the ULD.

[0101] - Path 2 (P = 0.30): The luggage is wrongly sent by OpB1 to the luggage pile of another remote stand flight in GH_B.

[0102] - Path 3 (P = 0.10): C2 receives the luggage but fails to record it due to a malfunction of the scanning device.

[0103] S3.6 GNN - Pay Intelligent Decision on Compensation Ratio: For Path 1: Construct the interaction graph of GH_B and C2. Node features: GH_B (high S2 score, high path participation, similar historical mistakes), C2 (medium S2 score, medium path participation, failure to fulfill the final confirmation responsibility). GNN - Pay output: GH_B: 60%, C2: 40%. For Path 2: Construct the interaction graph of GH_B (the main responsible party). GNN - Pay output: GH_B: 90%, C1: 10% (as the originator and signatory party, assume partial joint liability depending on the contract). Final comprehensive compensation suggestion: (60% * 0.60 + 90% * 0.30 + C2 _ Path3_Share * 0.10) for GH_B (for the visualization effect example, see the attachment Figure 3 as shown).

[0104] Step S4: Interactive Visualization, Decision Support and Closed - loop Feedback for Continuous Optimization S4.1 Visualization Interface: The claims adjuster can see the complete known trajectory of LH789XYZ in STKG - Bag on the interface. The transfer area of Airport B is highlighted as the breakpoint area. The paths 1 and 2 generated by DRL are superimposed on the simulated layout diagram of Airport B with different - colored dotted lines, and their respective probabilities and the compensation plans of GNN - Pay are shown beside. Clicking on the breakpoint area can view the M 2 analysis details of ARD - Net.

[0105] S4.2 Human - Machine Collaboration: The claims adjuster combines the real - time feedback of GH_B and C2 (such as CCTV investigation, manual search results). If Path 1 is confirmed (the luggage is found in the F2 container area), then confirm the path and the corresponding compensation plan. If CCTV shows that OpB1 did send the luggage to a wrong flight area (Path 2), then select Path 2.

[0106] S4.3 Continuous Learning: The confirmed cases (such as the final compensation result of Path 1 and GH_B: 60%, C2: 40%) are added to the training set. M2 When ARD-Net encounters a similar situation like "insufficient transfer time + missing C2 record" next time, it will be more inclined to locate the joint responsibility of GH_B and C2. The DRL agent will learn that in such a scenario, the probability of luggage being left behind in the container area is higher. GNN-Pay will fine-tune its logic for allocating compensation ratios among multiple parties based on more cases. In STKG-Bag, the weight of the pattern "insufficient transfer time at Airport B leads to a handover error from GH_B to C2" increases.

[0107] To verify the effectiveness of the method of the present invention, a simulation environment for airport luggage transportation and claims settlement was constructed, and simulation experiments were carried out. The method of the present invention was compared with traditional manual processing methods and automated systems based on simple rules in terms of performance.

[0108] The specific simulation environment of this method includes: It includes a network of 3 hub airports and 10 regional airports, simulating transfer scenarios of different complexities.

[0109] For the luggage data, 100,000 luggage transportation records were generated, including 5,000 abnormal luggage events (including various types such as loss, damage, misrouting, and delay). The introduction of abnormal events is based on the probability distribution statistically obtained from historical industry data. The luggage data includes tag numbers, PNRs, flight information, and multi-point scanning records (some records are missing or incorrect according to a set probability).

[0110] The operations of 5 major carriers and multiple ground service agents were set with their respective operation specifications, minimum connection times (MCTs), and historical operation error rates.

[0111] Based on the generated simulation data, the knowledge graph STKG-Bag was dynamically constructed for M 2 training of ARD-Net, DRL-CRQ, and GNN-Pay models.

[0112] The comparison methods and evaluation parameters are specifically as follows: Traditional manual method: Simulate processes such as manual record checking, phone communication, and email exchanges. The processing time and service quality are set based on the industry average.

[0113] System based on simple rules: An automated system based on predefined IF-THEN rules (for example, "if the A scanning point is missing and the B scanning point is normal, then the responsibility may lie between A and B").

[0114] Evaluation metrics: Responsibility determination accuracy, average claims settlement processing duration, adoption rate of automatically generated compensation plans, estimation of the improvement of passenger NPS (Net Promoter Score), and estimation of the reduction of disputes across airlines.

[0115] To visually display the comparison effect of the simulation results, please refer to the appendix Figure 4 , and at the same time, record the simulation data of the key performance indicators in the following table. The data in this table is consistent with the appendix Figure 4 content:

[0116] From the simulation results, it can be concluded that the method for monitoring the abnormal claim settlement process of airport luggage based on artificial intelligence proposed by the present invention has significantly improved the accuracy of liability determination, significantly shortened the average claim settlement processing time, and has a high adoption rate of automatically generated compensation plans, and is significantly superior to traditional manual methods and systems based on simple rules in various key performance indicators.

[0117] Through the above embodiments and simulation results, it shows that the method of the present invention can analyze, reason, quantify and provide decision support for complex and data-incomplete luggage abnormal events through a series of highly intelligent models, and finally achieve efficient, accurate and fair claim settlement processing.

[0118] In summary, the present invention realizes the automation and intelligence of luggage abnormal claim settlement by integrating dynamic knowledge graph, multi-modal deep learning and reinforcement learning technologies. The specific advantages are at least as follows: 1. By constructing a unified STKG-Bag and providing transparent decision-making basis, it helps to break data islands, promote trust and efficient cooperation among airlines, airports and ground services, and make full use of operation data.

[0119] 2. When data is missing or in conflict, the DRL-CRQ engine can intelligently deduce the most likely luggage path based on constraints and historical experience, provide key evidence for liability attribution, and effectively solve the "black box" problem.

[0120] 3. Through M2ARD-Net multi-modal information fusion and hierarchical attention, it can identify the real breakpoints of the liability chain from complex associations and subtle signals, and combine with interpretability analysis to improve the accuracy of liability determination.

[0121] 4. The GNN-Pay model can learn and recommend the multi-party compensation ratio based on the complex liability interaction graph, taking into account historical cases, industry rules and the specific situation of current events, and improving the fairness and negotiation efficiency of compensation.

[0122] 5. The automated and intelligent process shortens the investigation and decision-making time, reduces manual intervention, accelerates compensation, thereby improving the passenger experience and reducing operating costs.

[0123] The above are only the preferred embodiments of the present invention, and do not limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation made by using the content of the specification and drawings of the present invention, or directly or indirectly applied in other related technical fields, shall be similarly included in the patent protection scope of the present invention.

Claims

1. A method for monitoring the abnormal claim settlement process of airport luggage based on artificial intelligence, characterized in that, It includes the following steps: Real-time collect and fuse baggage transportation data from multi-source heterogeneous systems, and construct and dynamically maintain the spatio-temporal knowledge graph STKG-Bag of baggage transportation through knowledge extraction, entity alignment, and relationship inference. The STKG-Bag contains entities and spatio-temporal association relationships between entities. The entities include baggage, passengers, flights, carriers, operators, geographical locations, time points, operation events, and abnormal events; when a baggage abnormal event is triggered, extract the spatio-temporal sub-graph and operation event sequence related to the abnormal baggage from the STKG-Bag, and input the spatio-temporal sub-graph and operation event sequence into the pre-trained multi-modal attention fusion deep network M 2 ARD-Net. The M 2 ARD-Net obtains the analysis result of the breakpoint of the responsibility chain by encoding the graph structure context, the temporal event sequence, and the prior knowledge of operation specifications and abnormal patterns, and using multi-modal feature fusion and hierarchical attention mechanism; If the result of the responsibility chain breakpoint analysis shows that the actual trajectory of the luggage is unclear in a specific area, start the DRL-CRQ process of dynamic trajectory reconstruction and multi-party liability quantification based on deep reinforcement learning to obtain the reconstructed path and the compensation plan; The luggage trajectory, responsibility chain breakpoint analysis results, reconstruction path, and compensation plan in STKG-Bag are interactively displayed through an integrated multi-dimensional visualization interface. After receiving the final claim settlement plan input by the user, continuous incremental learning and iterative optimization are performed on the STKG-Bag, M 2 ARD-Net model, DRL-CRQ agent, and GNN-Pay model using the final claim settlement plan.

2. The method according to claim 1, wherein Knowledge extraction from luggage transportation data includes extracting entities and events from unstructured text using a pre-trained language model based on Transformer combined with conditional random fields or pointer networks, and relationship inference includes using graph neural networks for link prediction.

3. The method according to claim 1, characterized in that, The said M 2 The spatio-temporal context encoding module STCE-Module in ARD-Net includes: a graph spectral context encoder GSC of a graph attention network GAT for encoding the topological structure of the luggage transportation sub-graph and the dependencies between nodes; a temporal event sequence encoder TESE composed of a bidirectional gated recurrent unit Bi-GRU with self-attention mechanism for capturing the long-term dependencies and temporal dynamics of the operation event sequence; and a prior knowledge encoder PKE-Module for encoding operation specifications and historical anomaly patterns as prior knowledge and aligning them with the current event features.

4. The method according to claim 3, characterized in that, The M 2 ulti-modal Feature Fusion and Hierarchical Attention Module M of ARD-Net 2 The FA-Module uses a gated fusion unit to perform early fusion on the multi-modal features from the STCE-Module, then realizes in-depth interactive learning between features through a cross-modal Transformer encoder, and applies a hierarchical fine-grained attention mechanism to locate anomalies from macroscopic sections to microscopic events.

5. The method according to claim 1, wherein The said M 2 The responsibility attribution and breakpoint confidence evaluation module RAD-Module of ARD-Net uses a multi-task learning head to predict the breakpoint position probability, abnormal cause code, and preliminary tendency score of the responsible entity in parallel, and combines an interpretability method to highlight the input features that contribute greatly to the decision-making.

6. The method according to claim 1, wherein The starting of the DRL-CRQ process of dynamic trajectory reconstruction and multi-party liability quantification based on deep reinforcement learning to obtain the reconstructed path and the compensation plan specifically includes: Construct a dynamic environment for simulating luggage transfer based on STKG-Bag, operation rules, and breakpoint information; Through the deep reinforcement learning agent, use the Actor-Critic framework and combine Monte Carlo tree search to assist in exploration for policy learning in the environment, and reconstruct N most likely luggage movement paths and their probabilities according to the finely designed reward function; For each high-probability reconstructed path, construct an interaction graph of liability-related parties, and use the graph neural network compensation ratio allocation model GNN-Pay to quantify the compensation liability of each related party, and output an adaptive multi-party compensation ratio allocation plan; Among them, the reward function of the deep reinforcement learning agent comprehensively considers path coherence, time consistency, interpretability of anomaly recognition, satisfaction with operation rules and physical constraints, consistency with historical data patterns, and path complexity.

7. The method according to claim 6, wherein The input of the graph neural network compensation ratio allocation model GNN-Pay is the interaction graph of liability-related parties constructed based on the high-probability reconstructed path, where the node features include the preliminary liability inclination score of the related party, participation in the reconstructed path, historical operation error rate, and compliance with norms. The model outputs the compensation percentage of each related party and is trained by supervised learning of historical cases and industry rule constraints.

8. The method according to claim 1, characterized in that, The continuous incremental learning and iterative optimization include adopting an active learning strategy to select high-value samples for manual annotation, and using newly confirmed cases to evolve and update the knowledge graph, and performing periodic or event-triggered incremental training and parameter fine-tuning on the M 2 ARD-Net, DRL-CRQ agent, and GNN-Pay model.

9. An artificial intelligence-based monitoring system for the abnormal claim settlement process of airport luggage, characterized in that, It includes a memory and a processor. The memory stores a computer program, and when the processor executes the computer program, it implements the method for monitoring the abnormal claim settlement process of airport luggage based on artificial intelligence according to any one of claims 1 to 8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the method for monitoring the abnormal claim settlement process of airport luggage based on artificial intelligence according to any one of claims 1 to 8.

Images