An airport luggage abnormality claim process monitoring method based on artificial intelligence

Abstract

The application discloses an airport luggage abnormality claim process monitoring method based on artificial intelligence. The method first constructs and dynamically maintains a multi-dimensional luggage transportation space-time knowledge graph STKG-Bag. Secondly, a multi-modal attention fusion deep network M 2 ARD-Net intelligently analyzes and locates the responsibility chain of luggage abnormal events. Thirdly, when the luggage trajectory is ambiguous, a dynamic trajectory reconstruction and multi-party responsibility quantification process DRL-CRQ based on deep reinforcement learning is used to intelligently reconstruct the luggage path and adaptively optimize the multi-party compensation ratio using a graph neural network GNN-Pay. Finally, an interactive visualization interface is used to support decision-making and realize closed-loop feedback to continuously optimize each AI model and knowledge graph. The application integrates dynamic knowledge graph, multi-modal deep learning and reinforcement learning technology to realize the automation, precision, efficiency and fairness of luggage abnormality claim, and significantly improves the processing efficiency and passenger experience.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of airport operation management, and particularly relates to an airport luggage abnormal claim process monitoring method based on artificial intelligence. BACKGROUND

[0002] Airport luggage transportation is a core link of the civil aviation industry, directly affecting passenger experience and airline operation efficiency, and the accuracy and efficiency of its abnormal claim process are crucial to improving service quality. With the complication of global air transportation networks, luggage abnormal events occur frequently, involving multiple carriers and complex handover links, and the definition of responsibility for luggage abnormalities (such as loss, damage, delay) becomes exceptionally difficult. The traditional claim process highly depends on manual intervention, which is not only inefficient and costly, but also prone to passenger dissatisfaction and disputes between airlines due to information asymmetry, evidence chain breakage, and subjective judgment differences.

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

[0004] The development of deep learning and knowledge graph technology provides a new opportunity to solve such complex problems, but how to build a comprehensive solution that can adapt to the dynamic characteristics of aviation luggage 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

[0005] The present application aims to overcome the technical problems of fuzzy responsibility determination, lengthy processing period, strong subjectivity in decision-making, difficulty in cross-subject collaboration, and insufficient intelligence in existing airport luggage abnormal claim processes. A comprehensive monitoring and intelligent decision-making method is provided, which integrates dynamic knowledge graph construction, multi-modal deep learning driven responsibility boundary identification, deep reinforcement learning based fuzzy trajectory intelligent reasoning, and self-adaptive compensation scheme generation, to realize the automation, precision, efficiency, and fairness of luggage abnormal event processing.

[0006] The present application discloses an airport luggage abnormal claim process monitoring method based on artificial intelligence, comprising the following steps:

[0007] Real-time collection and fusion of baggage transportation data from multiple source heterogeneous systems, through knowledge extraction, entity alignment and relationship inference, construction and dynamic maintenance of baggage transportation spatio-temporal knowledge graph STKG-Bag, the STKG-Bag contains entities and spatio-temporal correlation between entities, the entities include baggage, passenger, flight, carrier, operator, geographic location, time point, operation event and abnormal event;

[0008] When the baggage abnormal event is triggered, the spatio-temporal subgraph and operation event sequence related to the abnormal baggage are extracted from the STKG-Bag, and the spatio-temporal subgraph and operation event sequence are input into the pre-trained multi-modal attention fusion deep network M 2 ARD-Net, the M 2 ARD-Net encodes the graph structure context, time sequence event sequence and operation specification and abnormal mode prior knowledge, and uses multi-modal feature fusion and hierarchical attention mechanism to obtain the responsibility chain breakpoint analysis result.

[0009] If the responsibility chain breakpoint analysis result shows that the actual trajectory of the baggage is unclear in a specific area, a deep reinforcement learning-based dynamic trajectory reconstruction and multi-party responsibility quantification DRL-CRQ process is started to obtain a reconstructed path and compensation scheme.

[0010] The baggage trajectory in the STKG-Bag, the responsibility chain breakpoint analysis result, the reconstructed path and the compensation scheme are interactively displayed through an integrated multi-dimensional visualization interface, and after receiving the final claim settlement scheme input by the user, the STKG-Bag, the M 2 ARD-Net model, DRL-CRQ agent and GNN-Pay model are continuously incrementally learned and iteratively optimized.

[0011] Preferably, the knowledge extraction of the baggage transportation data includes extracting entities and events from unstructured text using a Transformer-based pre-trained language model combined with a conditional random field or a pointer network, and the relationship inference includes link prediction using a graph neural network.

[0012] Preferably, the M 2The spatio-temporal context encoding module STCE-Module in the ARD-Net comprises: a graph structure context encoder GSC of a graph attention network GAT for encoding the topology of the baggage transportation subgraph and the dependencies between nodes; a time series event sequence encoder TESE constituted by a bidirectional gated recurrent unit Bi-GRU with a self-attention mechanism for capturing long-term dependencies and time dynamics of the operation event sequence; and a prior knowledge encoder PKE-Module for encoding the operation specification and historical abnormal patterns as prior knowledge and aligning them with the current event features.

[0013] Preferably, the M 2 The multi-modal feature fusion and hierarchical attention module M 2 The FA-Module adopts a gated fusion unit to perform early fusion of the multi-modal features from the STCE-Module, then realizes deep interaction learning between the features through a cross-modal Transformer encoder, and applies a hierarchical fine-grained attention mechanism to locate the anomaly from a macro section to a micro event.

[0014] Preferably, the M 2 The responsibility attribution and breakpoint confidence evaluation module RAD-Module of the ARD-Net adopts a multi-task learning head to predict the breakpoint position probability, abnormal reason code and responsibility subject preliminary inclination score in parallel, and highlights the input features that contribute greatly to the decision-making in combination with an explainability method.

[0015] Preferably, the starting of the dynamic trajectory reconstruction and multi-party responsibility quantification DRL-CRQ process based on deep reinforcement learning obtains a reconstructed path and a compensation scheme, and specifically comprises:

[0016] A dynamic environment simulating the movement of baggage is constructed based on the STKG-Bag, operation rules and breakpoint information;

[0017] A deep reinforcement learning agent adopts an Actor-Critic framework and combines a Monte Carlo tree search to assist exploration in the environment to perform policy learning, and reconstructs N most probable baggage movement paths and their probabilities according to a fine-tuned reward function;

[0018] For each high-probability reconstructed path, a responsibility-related party interaction graph is constructed, and a graph neural network pay-off proportion allocation model GNN-Pay is used to quantify the compensation responsibility of each related party, and an adaptive multi-party compensation proportion allocation scheme is output;

[0019] The reward function of the deep reinforcement learning intelligent agent comprehensively considers path continuity, time consistency, explanation of identifying abnormalities, satisfaction of operation rules and physical constraints, degree of conformity with historical data patterns and path complexity.

[0020] Preferably, the graph neural network pay ratio allocation model GNN-Pay inputs are liability-related party interaction graphs constructed based on high-probability reconstructed paths, wherein the node features include the preliminary liability inclination of the related party, the participation of the related party on the reconstructed path, the historical operation failure rate and the standard compliance degree, the model outputs the pay percentage of each related party, and the model is trained through supervised learning of historical cases and industry rule constraints.

[0021] Preferably, the continuous incremental learning and iterative optimization include selecting high-value samples for manual annotation by using an active learning strategy, and evolving and updating the knowledge graph by using newly confirmed cases, and updating the M 2 The ARD-Net, the DRL-CRQ intelligent agent and the GNN-Pay model are periodically or event-triggered for incremental training and parameter fine-tuning.

[0022] The application further discloses an airport luggage abnormality claim settlement process monitoring system based on artificial intelligence, which comprises a memory and a processor, and the memory stores a computer program.

[0023] The application further discloses a computer readable storage medium, which stores a computer program, and the computer program is characterized in that the computer program is executed by a processor to realize the airport luggage abnormality claim settlement process monitoring method based on artificial intelligence.

[0024] Compared with the prior art, the application integrates a dynamic knowledge graph deeply, constructs a unified STKG-Bag, helps to break data islands, and promotes trust and efficient cooperation among airlines, airports and ground services; when data is missing or conflicting, the DRL-CRQ engine can intelligently deduce the most possible luggage path based on constraints and historical experience, and provide key evidence for liability attribution; the M 2 ARD-Net can identify the real breakpoints of the responsibility chain from complex correlations and subtle signals through multi-modal information fusion and hierarchical attention, and improves the accuracy of responsibility determination. BRIEF DESCRIPTION OF DRAWINGS

[0025] Figure 1 is a multi-modal attention fusion deep network M 2 ARD-Net is used for intelligent analysis and breakpoint positioning of the responsibility chain.

[0026] Figure 2is the overall flowchart of the airport luggage abnormality claim process monitoring and intelligent decision method based on artificial intelligence in the embodiment of the application.

[0027] Figure 3 is the visual effect schematic diagram of the STKG-Bag subgraph, responsibility chain breakpoint highlighting, multiple fuzzy tracks of DRL reconstruction, and GNN-Pay compensation scheme suggestion of a complex multi-carrier transfer luggage abnormality case in the embodiment of the application.

[0028] Figure 4 is the simulation comparison bar chart of the method of the application and the traditional manual method and the method based on a simple rule system in key performance indicators (responsibility determination accuracy, average processing time, compensation scheme adoption rate, and passenger satisfaction index). DETAILED DESCRIPTION

[0029] In order to make the purpose, technical scheme and advantages of the application clearer, the application will be described in detail below with reference to the drawings and specific embodiments.

[0030] The embodiment discloses an airport luggage abnormality claim process monitoring method based on artificial intelligence (the overall process thereof is shown in the accompanying Figure 2 , and specifically can include:

[0031] Step S1: Construct and dynamically maintain a multi-dimensional luggage transportation space-time knowledge graph (STKG-Bag)

[0032] S1.1 Real-time acquisition and fusion engine of multi-source heterogeneous data: design an extensible data access and fusion framework to acquire multi-modal data from global distribution systems (GDS), airline departure control systems (DCS), luggage source messages (BSM), luggage handling system (BHS) logs, luggage reconfirmation system (BRS) scanning records, flight information display systems (FIDS), world tracking systems (WorldTracer), airport operation databases (AODB), ground agent operation platforms, and even authorized Internet of Things sensors (such as low-power wide-area network trackers on luggage tags) in real time or quasi-real time. The data includes: luggage tag information (IATA standard message), passenger itinerary (PNR), flight schedule and status, luggage scanning timestamp and location (accurate to sorting port, apron number, cargo hold), operator ID, electronic image of handover documents, abnormal report text, etc.

[0033] S1.2 STKG-Bag and Schema Design: Define a refined baggage transportation spatio-temporal knowledge graph ontology, including core entity types: Baggage, Passenger, FlightInstance, Segment, Carrier, Airport, Operator, GroundHandler, HandlingDevice, GeoLocation, Timestamp, HandlingEvent, ExceptionEvent, etc. Define rich attributes for each entity type (e.g., Baggage.tag_number, FlightInstance.actual_departure_time, HandlingEvent.operation_type, ExceptionEvent.report_reason_code). Define complex semantic relationships: e.g., Baggage.transport_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, causality, co-occurrence relationships, etc.

[0034] S1.3 Knowledge Extraction and Graph Instantiation:

[0035] S1.3.1 Structured Data Mapping: Directly map structured data from databases, API interfaces, etc. to entities and relationships in STKG-Bag.

[0036] S1.3.2 Semi-structured / Unstructured Data Extraction: Use pre-trained language models based on Transformers (e.g., BERT, ERNIE) combined with Conditional Random Fields (CRF) or Pointer Networks to extract key entities (e.g., specific described damage locations, passenger-mentioned handover details) and events from baggage exception report texts, operation notes, and handover document pictures (after OCR).

[0037] S1.3.3 Entity alignment and disambiguation: Handle the alignment problem of the same entity from different systems (e.g., different representations of the same flight in different systems) by using graph embedding-based entity alignment algorithms or rule-based disambiguation strategies. First, define the core set of identifying attributes for different types of entities. For example, for "flight", use "flight number + planned departure date + departure airport three-letter code + destination airport three-letter code" as features; for "passenger", use "(hashed) ID number + last name" or "frequent flyer number" as features. Use exact matching or edit distance (e.g., Jaro-Winkler algorithm, set a similarity threshold, e.g., 0.9 or above) to generate preliminary candidate entity pairs.

[0038] Second, use the inter-entity relationships already constructed in STKG-Bag to verify and adjust the weights. For example, if two preliminary matched "passengers" are associated with "luggage" with the same "luggage tag number", the confidence of aligning these two "passengers" is significantly improved. Conversely, if their associated flight information conflicts, the confidence is reduced.

[0039] Third, apply specific rules in the luggage transportation field for forced alignment or exclusion. For example, according to IATA Resolution 753, luggage records with the same 10-digit license plate number should be highly prioritized to align as the same piece of luggage, even if other information is slightly different.

[0040] Finally, for low-confidence alignment results or conflicting alignments, a manual review mechanism can be introduced, and the review results can be fed back to the alignment model for learning, achieving iterative optimization.

[0041] 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 priority; (4) iteration and manual review.

[0042] S1.3.4 Relationship inference and link prediction: Use graph neural networks (GNN) such as GraphSAGE or RotatE models to train on the constructed graph fragments to infer potential missing relationships or predict future possible links (e.g., predict which ULD the luggage is most likely to be loaded into).

[0043] S1.4 Dynamic updating and version control of graph: Ensure that STKG-Bag can reflect the latest status of luggage and changes in the transportation environment in near real-time. Introduce event traceability mechanism and version control function of graph database to track historical state and perform backtracking analysis.

[0044] Step S2: Responsibility chain intelligent dissection and breakpoint positioning based on multi-modal attentive fusion deep network (M 2 ARD-Net)

[0045] When a baggage abnormality event (such as Baggage Irregularity Report, PIR) is triggered, all historical operation event sequences, involved carriers, handover points, timestamps, operators, etc. related to the abnormal baggage are extracted from STKG-Bag to form a spatio-temporal subgraph. A multi-modal attentive fusion deep network (M 2 ARD-Net, Multi-modal Attentive Responsibility Dissection Network) is constructed, and its detailed network architecture is shown in the accompanying Figure 1 for accurate positioning of the responsibility chain breakpoint and preliminary assessment of the abnormal contribution of each link.

[0046] S2.1 Spatio-temporal context encoding module (STCE-Module):

[0047] S2.1.1 Graph structure context encoder (GSCE): A multi-layer graph attention network (GAT) is used to process and extract the baggage transportation subgraph. GAT can learn the importance of nodes (such as carriers, handover points, operators) in the graph structure and capture their complex interdependent relationships. The initial features of the nodes can be composed of their embedding vectors or attributes in STKG-Bag. The context-aware representation of each node is output.

[0048] S2.1.2 Temporal event sequence encoder (TESE): The time-ordered baggage operation event sequence (each event contains features such as time, location, operation type, operator ID, etc.) is encoded. A bidirectional gated recurrent unit (Bi-GRU with Self-Attention) network with self-attention mechanism is used. Bi-GRU captures the long-term dependencies and temporal dynamics between events, and the self-attention mechanism helps the model focus on a few key events in the sequence that are most critical to the abnormality.

[0049] S2.1.3 Operation specification and abnormal pattern prior knowledge encoder (PKE-Module):

[0050] a. Operation specification library embedding: The key rules in the airline's standard operating procedures (SOPs) and the IATA baggage handling manual (BHM) are structured (such as rule graphs or vector embeddings) to form an operation specification knowledge base.

[0051] b. Historical anomaly pattern library: From historical solved baggage abnormality cases, typical abnormality patterns are extracted by clustering or pattern mining algorithms (e.g. Apriori combined with sequential pattern mining) (e.g. "when the handover time of a certain route A-B is less than 30 minutes, the baggage misdelivery rate is significantly increased").

[0052] c. Priori knowledge fusion: Using a small attention network, the current event sequence features are matched and aligned with the relevant entries in the operation specification library and the historical anomaly pattern library, generating a "specification deviation degree" or "historical pattern coincidence degree" feature vector. For example, the rule "when the transit time is less than the minimum connection time (MCT) specified by the airport, the fast transfer procedure should be started" can be represented as a structured object containing the condition (transit time < MCT), context (specific airport / carrier) and expected action (start fast transfer), and its text description or structured features are encoded into dense vectors by pre-training language models (such as BERT). For example, a mined abnormality pattern may be: "event sequence: [flight A arrives at transfer station X] -> [baggage is unloaded and scanned at X station] -> [long time (e.g. more than 1.5 times the average transit time at X station) without subsequent loading scan] -> [passenger reports luggage loss at destination Y]". These extracted patterns are also converted into feature vectors or graph patterns for similarity matching with the current event sequence.

[0053] S2.2 Multi-modal feature fusion and hierarchical attention module (M 2 FA-Module):

[0054] S2.2.1 Early fusion and gating mechanism: The graph structure context representation output by GSCE, the time sequence event representation output by TESE, and the priori knowledge representation output by PKE are preliminarily fused through a learnable gating mechanism (Gated Fusion Unit), which can dynamically adjust the weights of different modal information.

[0055] This gating unit dynamically weights the relative importance of different information in the baggage claim scene, for example, in the scenario where the baggage trajectory record is complete but the responsibility is ambiguous, the operation specification compliance analysis provided by the PKE module may have a higher weight; while in the case of missing key data in the baggage trajectory, the historical behavior patterns of related entities (such as carriers, operators) provided by GSCE and the event sequence context provided by TESE may be more important.

[0056] S2.2.2 Cross-modal Interaction Attention Layer: Based on the fused features, a multi-layer cross-modal Transformer encoder is applied. This encoder enables deep interaction between features from different sources (e.g., temporal and spatial / structural dimensions) through self-attention and feed-forward network layers, learning complex non-linear relationships between them and identifying subtle abnormal signals hidden in multi-modal data.

[0057] S2.2.3 Hierarchical Fine-grained Attention: A hierarchical attention mechanism is designed to first allocate attention at a macro level (e.g., flight segments, major handover points) to identify the most suspicious responsibility segments, and then further refine attention within the identified suspicious segments to specific individual operational events or time points, achieving coarse-to-fine abnormality localization.

[0058] S2.3 Responsibility Attributing and Breakpoint Confidence Assessment Module (RAD-Module):

[0059] S2.3.1 Multi-task Breakpoint Prediction Head: The fused final feature representation is fed into a multi-task learning head, which performs the following tasks in parallel:

[0060] a. Breakpoint Position Classification: For each operational event or handover link, the probability of it being a responsibility chain breakpoint is predicted (binary or multi-class classification if the breakpoint has types).

[0061] b. Exception cause code prediction: predict the most likely cause code of the break (e.g. IATA standard cause code: loading error, offloading error, documentation error, transfer mishandling, etc.). The exception cause code system adopted by the present invention is based on the IATA standard baggage mishandling cause code (e.g. the code in IATA AHM Chapter 7, such as 52-Transfer Mishandled, 73-Failure to Load at Transfer), and is further refined and expanded according to actual operation experience and data statistics, forming a hierarchical classification system. For example, “73-Failure to Load at Transfer” is further divided into “73.1-ULD not loaded as planned”, “73.2-baggage sorting error not delivered to the designated ULD”, “73.3-baggage has arrived at the ULD but not loaded due to lack of space / time”, etc. to more accurately locate the problem. A preliminary exception relevance or responsibility tendency score is given to each responsible subject involved in this link (e.g. carrier, ground service agent, airport operation department). The basis of the score mainly comes from the patterns learned by the model from the input features, for example, whether the last operation of the subject before the break conforms to the operation specification coded by the PKE module, the historical failure rate of the subject recorded in the STKG-Bag for similar scenarios, the actual control of the subject over the key control points in the current event chain (such as baggage handover), etc. The score result is usually normalized to the interval [0, 1], indicating the association strength or potential responsibility size of the subject to the current exception event.

[0062] c. Preliminary tendency score of responsibility subject: give a preliminary exception relevance score to each responsible subject involved in this link (carrier, ground service agent).

[0063] S2.3.2 Explainability analysis: Use explainable AI methods such as Integrated Gradients or LIME to highlight the most contributing parts in input features to breakpoint determination and cause prediction (e.g., a missing scan record, an abnormal time interval, an irregular operator signature), providing evidence for manual review. The output is a structured breakpoint report containing breakpoint location, confidence, predicted cause, preliminary responsible party tendency, and supporting evidence snippets. For example, when the model attributes the responsibility of a "baggage not loaded for subsequent flight" mainly to the ground service at the transfer airport, the explainability analysis may highlight "in the input event sequence, the ground service's operation time at the transfer baggage handover link is far beyond the average level" and "the prior knowledge of 'the ground service is short of manpower resources at this time period' matched by the PKE module", helping the manual review personnel understand the model's decision logic.

[0064] Step S3: Fuzzy trajectory intelligent reconstruction and multi-party responsibility quantification based on deep reinforcement learning and constraint programming (DRL-CRQ)

[0065] This step is activated when step S2 identifies that the responsibility chain has significant breakpoints (e.g., missing critical scan records, abnormal handover times), resulting in a fuzzy actual trajectory of the baggage in a certain area.

[0066] S3.1 Dynamic environment construction and state space definition:

[0067] S3.1.1 Environment: Based on the confirmed baggage trajectory fragments in STKG-Bag, airport facility layout (from AODB or GIS), flight schedule, carrier / ground service operation capability constraints (such as processing capacity per unit time, standard transfer time window), and breakpoint information from S2 (as "signals" or "disturbances" in the environment), a simulated dynamic baggage flow environment is constructed.

[0068] S3.1.2 State space (S): The state of the DRL agent includes: the current assumed baggage location (discrete or continuous), the current time, the known history of the baggage trajectory, the breakpoint information and confidence given by S2, the congestion index of the surrounding environment (if available), and the constraint satisfaction of the current path. Airport facility layout (topology graph represented in graph database or JSON format, nodes representing key processing units such as check-in island, security channel, baggage carousel, sorting slot, ULD group board area, and apron position, edges representing passable paths and their standard transit times, precision to specific functional areas or key equipment) operation capability constraints (for example, a certain sorting belt has a maximum processing capacity of N pieces per minute, and a certain ground service can handle M pieces of baggage per hour during peak transfer time, these parameters are used as resource limitations and behavior constraints in the simulated environment).

[0069] S3.1.3 Action Space (A): The actions of the agent include: selecting the next possible physical location (e.g., a certain sorting slot, a transit storage area, a ULD, a cargo bay), selecting the next logical operational state (e.g., "waiting for loading", "in transit", "unloaded"), or "maintain the current state and wait" (simulating a delay). The action space is dynamically generated according to the current location and airport operation rules. For example, if the current simulated baggage is located at the exit Z of the transit sorting system, and the SOP specifies that exit Z only handles baggage for flights to the R1 region, the agent's available actions will only include directing the baggage to the container area or conveyor associated with the R1 region flights, and will not generate actions to direct it to other regions.

[0070] S3.2 Policy Network and Value Network Design (Based on Actor-Critic Framework): Adopt Actor-Critic algorithms (e.g., improved versions of A3C, Soft Actor-Critic (SAC), or Proximal Policy Optimization (PPO)).

[0071] S3.2.1 Actor Network (Policy Network): Input the current state S, output the probability distribution of taking each possible action A in the current state (for discrete action space) or deterministic action (for continuous action space). The policy network itself can be a deep neural network containing convolutional layers (if the location is an image representation), recurrent layers (to handle temporal information in the state), and fully connected layers.

[0072] S3.2.2 Critic Network (Value Network): Input the current state S (sometimes also input the action A), output the value evaluation (Q value or V value) of the current state (or state-action pair), used to guide the learning of the Actor network.

[0073] S3.3 Reward Function Design: The reward function is the core of guiding the agent to reconstruct the most "reasonable" baggage trajectory, and its design goal is to make the reconstructed trajectory both explainable and consistent with physical and operational logic. The reward function R is the sum of multiple weighted sub-items:

[0074] 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.

[0075] Examples of specific quantification of each sub-item are as follows:

[0076] a. Path coherence reward R_coherence: If the reconstructed path segment can be smoothly connected with the known trajectory segment in STKG-Bag at both the start and end points in logical and physical locations (e.g., spatial distance less than threshold ε), then R_coherence is a large positive value (e.g., +20), otherwise 0 or negative.

[0077] b. Temporal consistency reward R_temporal: For each simulated operation step in the path, if its time consumption is consistent with the standard operation time or flight schedule (e.g., MCT), then a positive reward is given; if there is an unexplainable long stall or super-high speed movement, then a significant negative reward is given (e.g., -10 penalty if the time deviation is more than 50% of the standard operation time).

[0078] c. Abnormal explanation reward R_explanation: If the reconstructed path can reasonably explain the breakpoints or abnormal phenomena identified in step S2 (e.g., the path shows that the baggage entered an area where misrouting often occurs before the breakpoint, or the path conforms to a certain known abnormal pattern encoded in the PKE module), then R_explanation is a large positive value (e.g., +30).

[0079] d. Constraint satisfaction reward / punishment R_constraint: If the path strictly adheres to airport operation rules (e.g., the baggage does not enter the customs supervision area, etc.), physical constraints (e.g., the baggage cannot pass through walls), and baggage handling SOPs, then R_constraint is 0 or a small positive value; for each hard constraint violation, a large negative penalty is added (e.g., -100).

[0080] e. Historical data match reward R_historical_match: If the pattern of the reconstructed path (e.g., key processing node sequence, involved operator type combination) is highly consistent with the pattern of a historical baggage real trajectory or high-confidence reconstructed trajectory recorded in the STKG-Bag with similar initial conditions (e.g., same O&D, same carrier combination, similar abnormal type) (e.g., measured by trajectory edit distance or sequence embedding cosine similarity, similarity greater than threshold 0.8), then R_historical_match is a positive value, the higher the consistency, the greater the reward.

[0081] f. Path complexity penalty P_complexity: To avoid generating excessively long or unnecessarily complex detour paths, a slight negative penalty proportional to the path length / turning number is imposed on the reconstructed path length (number of operation steps) or turning number (number of direction changes) (e.g., -0.1 penalty for each additional step or turning).

[0082] The weight coefficients w_coh, w_time, w_exp, w_con, w_hist, and w_comp are determined through multiple rounds of experiments and parameter tuning in a simulated environment.

[0083] S3.4 Monte Carlo Tree Search (MCTS) assisted exploration and planning: At certain decision points, especially when the action space is large or the environment is highly uncertain, MCTS can be combined to assist the DRL agent in deeper exploration and planning. Specifically, each iteration of MCTS includes four stages of selection, expansion, simulation, and backtracking. In the selection and expansion stages, the action probability distribution output by the DRL policy network (Actor network) can be used as a prior strategy for MCTS tree search to guide node selection and new node expansion. In the simulation (Rollout) stage, the DRL policy network can be used to quickly execute to the end, or the DRL 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 amount of calculation. After MCTS completes a specified number of iterations, the access times or cumulative Q values of each child node (corresponding to all available actions in the current state) under the root node can be used to generate a more optimal action selection strategy (e.g., selecting the action with the most access times), which can be directly used to guide the agent to make decisions in the current simulation step, or used to generate high-quality (state-improved policy) sample pairs to further train and optimize the DRL policy network.

[0084] S3.5 Multi-path generation and probability evaluation:

[0085] The DRL agent may explore and evaluate multiple candidate reconstruction paths. The system records the N most likely paths and their corresponding cumulative rewards or probabilities given by the policy network.

[0086] S3.6 Responsibility quantification and pay ratio intelligent decision-making based on the set of reconstruction paths (GNN-Pay):

[0087] S3.6.1 Constructing responsibility-related party interaction graph: For each high-probability reconstruction path, identify all relevant responsible subjects (carriers, ground services, airport departments) on the path. These subjects are nodes, and their operational handovers or contractual relationships are edges, constructing a small interaction graph. Node features include: the preliminary responsibility inclination score assigned to this subject in S2, its participation in the current reconstruction path (such as operation duration, key node control), its historical operation failure rate (extracted from STKG-Bag), and whether it violates relevant operation specifications.

[0088] S3.6.2 Graph Neural Network Payoff Allocation Model (GNN-Pay): Design a graph neural network (e.g., Graph Isomorphism Network, GIN, or GAT with attention mechanism) that takes the above-mentioned liability-related interaction graph as input. The GNN can learn the patterns of responsibility allocation under complex interactions among multiple parties. The output layer of the model is a Softmax layer that directly outputs the percentage of compensation for each responsible party in relation to the current abnormal event. During training, the model can use historical resolved compensation cases (including the responsibility proportion of each party) as a supervision signal, and combine industry rules such as the IATA Multilateral Interline Baggage Compensation Agreement (MIPA) as a constraint or regularization term. Specifically, in the loss function of the GNN model, in addition to the standard supervised loss (such as cross-entropy loss or mean square error loss, used to fit the historical cases' compensation proportion), an additional regularization term is added to penalize the difference between the model's output compensation proportion and the responsibility division proportion suggested by MIPA rules in applicable scenarios. For example, if MIPA stipulates that the origin carrier bears X% responsibility and the transfer carrier bears Y% responsibility in a certain interline situation, 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 proportions of the origin and transfer carriers output by the GNN model, and λ_mipa is a hyperparameter that controls the strength of the regularization term. In this way, while learning the complex patterns of historical data, the GNN-Pay model can also ensure that its output results conform to industry-recognized fairness standards to some extent.

[0089] S3.6.3 Multi-path Result Fusion: If there are multiple high-probability reconstruction paths, the compensation proportions output by GNN-Pay for each path can be weighted and averaged (with weights being the probabilities or confidences of the paths) to obtain the final comprehensive compensation scheme recommendation.

[0090] Step S4: Interactive Visualization, Decision Support, and Closed-loop Feedback Continuous Optimization

[0091] S4.1 Multi-dimensional visualization and intelligent interpretation dashboard: The complete life cycle of the baggage in STKG-Bag (including the fuzzy part reconstructed by S3), M 2 The responsibility chain breakpoints identified by ARD-Net (highlighted with accompanying reasons), the multiple candidate trajectories generated by DRL-CRQ (with varying colors or transparency to indicate probability), 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 detailed analysis and simulating the impact of different compensation plans. A summary report generated in natural language is provided to explain the decision-making process and key rationales.

[0092] S4.2 Human-Machine Collaborative Decision-Making and Audit Trail: Claims specialists can make final decisions based on the system's intelligent analysis results. The system should allow specialists to intervene and correct key decisions (such as breakpoint confirmation, path selection, and compensation ratio adjustment). All manual operations and justifications should be recorded to form a complete audit trail.

[0093] S4.3 Feedback-based continuous learning and model evolution:

[0094] S4.3.1 Active Learning and Labeling: For judgments with low system confidence or cases with a large number of manual corrections, the active learning mechanism is triggered to request claims experts to perform high-quality labeling and expand the training data set.

[0095] S4.3.2 Model incremental training and fine-tuning: Regularly (or after accumulating a certain amount of feedback data) use newly confirmed cases (including manual correction results) to fine-tune the M 2 ARD-Net, DRL policy / value networks, and GNN-Pay models are incrementally trained or fine-tuned to adapt to changes in the operating environment, emerging abnormal patterns, and updates to compensation rules. Transfer learning techniques are used to effectively transfer knowledge learned from old data to new models.

[0096] S4.3.3 Knowledge graph evolution: Dynamically update the STKG-Bag ontology (such as adding new abnormal cause types) and instance data based on claim results and newly discovered patterns to ensure the freshness and accuracy of the knowledge graph.

[0097] To further illustrate the specific operation process of the method of the present invention and its application effect in actual scenarios, a typical complex multi-carrier transfer baggage anomaly case will be used for detailed description below.

[0098] This embodiment will demonstrate how the present invention gradually applies the aforementioned core steps, including the construction of a spatiotemporal knowledge graph for baggage transportation, intelligent analysis of the responsibility chain, intelligent reconstruction of fuzzy trajectories, quantification of multi-party responsibilities, and interactive decision support, to achieve efficient and accurate handling of baggage anomaly incidents.

[0099] Suppose a passenger's checked baggage (tag LH789XYZ) with the planned itinerary: City S1 (Airport A, Carrier C1) -> City S2 (Airport B, C1 and C2 code share and handoff here, Ground Handler GH_B responsible for actual transfer operation) -> City S3 (Airport C, Carrier C2). The passenger reports the baggage missing at Airport C.

[0100] Step S1: Construct and dynamically maintain multi-dimensional baggage transportation space-time knowledge graph (STKG-Bag)

[0101] S1.1 Multi-source heterogeneous data real-time collection and fusion engine: The system obtains from C1's DCS the record of LH789XYZ's check-in at Airport A (time T0, operator OpA1, weight, destination C), and BSM shows that the baggage has been loaded onto flight F1 (S1-S2). From BRS at Airport B, it is obtained that after flight F1 arrives, LH789XYZ is scanned in BHS (time T1, location SortPier_X). From GH_B's operation platform, a record is obtained showing that baggage LH789XYZ is assigned to flight F2 (carrier C2) to S3, and there is an electronic handover action (time T2, operator OpB1 signs to hand over to C2 for loading). From C2's DCS, no record of LH789XYZ's loading on flight F2 is found. WorldTracer has a PIR submitted by passenger P001 at Airport C, describing the characteristics of the baggage.

[0102] S1.2 STKG-Bag Ontology and Schema Design: Define a refined baggage transportation spatiotemporal knowledge graph ontology, containing core entity types: Baggage, Passenger, FlightInstance, Segment, Carrier, Airport, Operator, GroundHandler, HandlingDevice, GeoLocation, Timestamp, HandlingEvent, ExceptionEvent, etc. Define rich properties for each entity type (e.g., Baggage.tag_number, FlightInstance.actual_departure_time, HandlingEvent.operation_type, ExceptionEvent.reported_reason_code). Define complex semantic relationships: e.g., Baggage.transport_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, causality, co-occurrence, etc. Create entities according to the 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. Property filling: Baggage:LH789XYZ.destination = C, HandlingEvent:CheckIn_A.timestamp = T0, ExceptionEvent:PIR_C.reported_reason = "Not_Received_At_Destination".

[0103] Relationship Construction: 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.

[0104] S1.3 Knowledge Extraction and Graph Instantiation: The text "Baggage: LH789XYZ color = "Blue", Baggage:LH789XYZ.shell_type = "Hard", Baggage:LH789XYZ.distinguishing_mark = "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 transfer sheet, the key information is extracted after OCR recognition. According to the scheduled departure time T_F2_dep of flight F2 and the time T2 of TransferPrep_B_F2, the system finds that T_F2_dep-T2<minimum transit handling time (MCT), which may be a potential risk point marked by the rule engine or GNN link prediction module in the graph.

[0105] S1.4 Dynamic Update of Graph:

[0106] If C2 later finds LH789XYZ in a corner of airport B and supplements the scanned information, STKG-Bag will update the status and location of the baggage in real time.

[0107] Step S2: Based on M 2 Intelligent Analysis and Breakpoint Location of Responsibility Chain of ARD-Net

[0108] 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)" -> [C2 loading record missing] -> "C_PIR_Reported (P001)".

[0109] wherein the extracted event sequence is specified as follows:

[0110] Event "A_CheckIn (T0, C1)": This event indicates that the baggage LH789XYZ was recorded as checked-in and shipped by the agent or system of the carrier C1 at the origin airport A at the time point T0. In the STKG-Bag, this event node is associated with the baggage LH789XYZ, the passenger P001, the flight F1 (subsequent carrier flight), the carrier C1, the geographical location airport A, and the time point T0.

[0111] Event "A_Load_F1 (C1)": This event indicates that the baggage LH789XYZ was recorded as loaded onto the flight F1 operated by the carrier C1 at the origin airport A. In the STKG-Bag, this event node is associated with the baggage LH789XYZ, the flight F1, the carrier C1, and the corresponding time and location of the operation, and has a chronological relationship with the event "A_CheckIn (T0, C1)".

[0112] Event "B_Unload_F1 (GH_B for C1)": This event indicates that the baggage LH789XYZ was recorded as unloaded from the flight F1 after the flight F1 arrived at the transit airport B, and this operation was performed by the ground agent GH_B on behalf of the carrier C1. In the STKG-Bag, this event node is associated with the baggage LH789XYZ, the flight F1, the ground agent GH_B (as the operation performer, and indicates that it serves the carrier C1), the geographical location airport B, and the corresponding time and location of the operation.

[0113] Event "B_Sort (T1, GH_B)": This event indicates that the baggage LH789XYZ was sorted by the ground agent GH_B's operator or automated system at the time point T1 at the transit airport B. In the STKG-Bag, this event node is associated with the baggage LH789XYZ, the ground agent GH_B, the time point T1, and the specific sorting area within the airport B.

[0114] Event "B_TransferPrep_F2 (T2, OpB1 of GH_B for C2)": This event indicates that the baggage LH789XYZ was prepared for transfer to the subsequent flight F2 operated by the carrier C2 at time point T2 at transfer airport B by the operator OpB1 of the ground handling agent GH_B. In STKG-Bag, this event node is associated with the baggage LH789XYZ, the flight F2 (the target transfer flight), the operator OpB1 (affiliated to GH_B), the carrier C2 (as the subsequent carrier), the time point T2, and the corresponding operation location (e.g., the transfer baggage sorting area or the handover area).

[0115] Breakpoint "[C2 loading record missing]": This is a critical potential breakpoint in the responsibility chain. In STKG-Bag, the system did not query or receive a valid record from the carrier C2 that the baggage LH789XYZ was loaded onto the flight F2 (e.g., the corresponding BSM LDM message or DCS loading confirmation information is missing). This indicates that there is uncertainty in the transportation status of the baggage after the event "B_TransferPrep_F2 (T2, OpB1 of GH_B for C2)", which is the focus of the M²ARD-Net for responsibility analysis and breakpoint localization.

[0116] Event "C_PIR_Reported (P001)": This event indicates that the passenger P001 submitted a baggage irregularity report (PIR) at the final destination airport C due to not receiving the baggage LH789XYZ. In STKG-Bag, this event node (as an abnormal event entity) is associated with the baggage LH789XYZ, the passenger P001, the geographical location airport C, and the report submission time, and serves as the triggering event for the entire abnormal claim settlement process.

[0117] S2.1 STCE-Module: GSCE analyzes the handover relationship and historical cooperation records between C1, C2, and GH_B in this subgraph (if available in the graph). TESE processes the event sequence, focusing on the missing expected operation of C2 after T2. The PKE module matches the historical pattern of "GH_B's handover error rate to C2 rising during the tight transfer time at airport B" (assuming it exists), and notes the operation specification deviation of T_F2_dep - T2 < MCT.

[0118] S2.2 M 2FA-Module: The gated fusion unit balances graph structure information, temporal information, and prior knowledge. The cross-modal Transformer deeply integrates these features. For example, it discovered 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 transfer time" in the prior knowledge. Hierarchical attention first focuses on the "transfer link at Airport B" and then refines it to the specific operation of "GH_B handing over flight F2's baggage to C2."

[0119] S2.3 RAD-Module:

[0120] Multi-task header output:

[0121] - Breakpoint location: After HandlingEvent:TransferPrep_B_F2, during the loading phase of FlightInstance:F2, with a confidence level of 0.95.

[0122] - Abnormal reason code (forecast): IATA Code 52: Transfer Mishandled or IATA Code 73: Failure to Load at Transfer.

[0123] The initial tendency toward liability is: GH_B (operator, possibly due to time constraints) > C2 (failure to confirm receipt and loading). Explainability analysis highlights: C2's BSM loading message is missing, GH_B operator OpB1's sign-off record at T2 lacks a corresponding receipt confirmation from C2, and the objective fact of insufficient transit time.

[0124] Step S3: DRL-CRQ-based fuzzy trajectory intelligent reconstruction and multi-party responsibility quantification

[0125] Due to the missing loading record of C2, the baggage's trajectory from after T2 to before flight F2 took off is unclear.

[0126] S3.1 Dynamic Environment Construction: This environment simulates the GH_B operating area at Airport B, the loading area / port for Flight F2 at C2, and possible error flow areas. The status includes the baggage's current assumed location, time, remaining time until F2's departure, and breakpoint signals identified by S2.

[0127] S3.2 Policy and Value Network (SAC Algorithm): The Actor Network learns the strategy for moving luggage to the next location or operation state under different states. The Critic Network evaluates the quality of the strategy.

[0128] S3.3 Reward Function:

[0129] - If the path shows that the baggage is correctly delivered to C2 before F2 departure and there is a simulated loading action, give a high positive reward.

[0130] - If the path shows that the baggage is not delivered to C2 due to GH_B operation failure (e.g. sent to the wrong sorting port, forgotten in the temporary storage area), and the time coincides with the deadline for receiving F2, give an explanatory positive reward.- If the path violates some operation rules of airport B (e.g. entering a non-baggage area), give a large negative punishment.

[0131] - The path coincides with the historical pattern of “insufficient transit time leading to GH_B operation confusion”, give a reward.

[0132] S3.4 MCTS assistance: when there are multiple possible flow directions in the GH_B operation area, MCTS helps to explore the optimal few moving directions.

[0133] S3.5 Multi-path generation:

[0134] DRL generation example:

[0135] - Path 1 (P=0.60): the baggage is sent by OpB1 to the ULD assembly area of F2 flight at T2, but due to time pressure or the receiving personnel of C2 being busy, the baggage is missed at the edge of the assembly area and not loaded into the ULD.

[0136] - Path 2 (P=0.30): the baggage is sent by OpB1 to the baggage pile of another remote flight of GH_B.

[0137] - Path 3 (P=0.10): C2 receives the baggage but the scanning equipment fails to record.

[0138] S3.6 GNN-Pay compensation ratio intelligent decision: for path 1: build the interaction graph of GH_B, C2. Node features: GH_B (high S2 score, high path participation, similar historical failure), C2 (medium S2 score, medium path participation, not final responsibility). GNN-Pay output: GH_B: 60%, C2: 40%. For path 2: build the interaction graph of GH_B (mainly responsible party). GNN-Pay output: GH_B: 90%, C1: 10% (partially liable as the originator and signatory, subject to contract). Final comprehensive compensation recommendation: (60%*0.60 + 90%*0.30+C2 _ Path3_Share * 0.10) for GH_B (its visualization example is shown in FIG. 6). Figure 3

[0139] Step S4: Interactive visualization, decision support and closed-loop feedback for continuous optimization

[0140] S4.1 Visualization interface: The claim handler sees the complete known trajectory of LH789XYZ in STKG-Bag, the transit area in airport B is highlighted as the breakpoint area, path 1 and path 2 generated by DRL are superimposed on the simulated layout of airport B as dashed lines in different colors, with the respective probabilities and the payout scheme of GNN-Pay displayed next to them. Clicking on the breakpoint area allows viewing of M 2 Analysis details of ARD-Net.

[0141] S4.2 Human-machine collaboration: The claim handler combines real-time feedback from GH_B and C2 (such as CCTV investigation, manual search results), and if path 1 is confirmed (the luggage is found in F2 packing area), the path and corresponding payout scheme are confirmed. If CCTV shows that OpB1 indeed delivered the luggage to an incorrect flight area (path 2), path 2 is selected.

[0142] S4.3 Continuous learning: Confirmed cases (such as path 1 and GH_B: 60%, C2: 40% final payout result) are added to the training set. M 2 Next time ARD-Net encounters a similar situation of "insufficient transit time + missing C2 records", it will be more inclined to locate the joint responsibility of GH_B and C2. The DRL agent will learn that in this scenario, the probability of luggage missing in the packing area is higher. GNN-Pay will fine-tune its payout proportion allocation logic among multiple parties based on more cases. In STKG-Bag, the pattern weight of "insufficient transit time in airport B leading to GH_B handover failure to C2" increases.

[0143] To verify the effectiveness of the method, a simulated airport baggage transportation and claim handling environment is constructed, and a simulation experiment is conducted to compare the performance of the method with traditional manual processing methods and automated systems based on simple rules.

[0144] The specific simulation environment of the method includes:

[0145] It contains a network of 3 hub airports and 10 regional airports, simulating transit scenarios of different complexity.

[0146] The luggage data generates 100,000 luggage transportation records, including 5,000 abnormal luggage events (including loss, damage, misdelivery, delay, etc.), and the introduction of abnormal events is based on the probability distribution of historical industry data statistics. Luggage data includes tag number, PNR, flight information, multi-point scanning records (some records are missing or incorrect according to the set probability).

[0147] 5 major carriers and multiple ground service agents operate, setting their own operating specifications, minimum transit time (MCT) and historical operating failure rates.

[0148] Based on the generated simulation data, the knowledge graph STKG-Bag is dynamically constructed, and M 2 ARD-Net, DRL-CRQ, and GNN-Pay model training.

[0149] The comparison method and evaluation parameters are as follows:

[0150] Traditional manual method: simulate manual record checking, telephone communication, email exchange and other processes, and the processing time and service quality are set based on the industry average level.

[0151] Simple rule-based system: an automated system based on predefined IF-THEN rules (for example, "if A scan point is missing and B scan point is normal, then the responsibility may be between A-B").

[0152] Evaluation indicators: responsibility determination accuracy, average claim processing time, adoption rate of automatically generated compensation plan, estimated passenger NPS (Net Promoter Score) improvement, and estimated cross-airline dispute rate reduction.

[0153] For intuitive display of the simulation results, see the attached Figure 4 At the same time, the simulation data of the key performance indicators are recorded in the following table, and the data in this table are consistent with the attached Figure 4 content:

[0154]

[0155] From the simulation results, it can be concluded that the airport baggage abnormal claim process monitoring method based on artificial intelligence proposed by the present application has significantly improved responsibility determination accuracy, significantly shortened average claim processing time, and high adoption rate of automatically generated compensation plan, which is significantly superior to traditional manual methods and simple rule-based systems in all key performance indicators.

[0156] Through the above examples and simulation results, the method of the present application can analyze, reason, quantify and decision support complex and incomplete data baggage abnormal events through a series of highly intelligent models, and finally achieve efficient, accurate and fair claim processing.

[0157] In summary, the present application realizes the automation and intelligent method of baggage abnormal claim by fusing dynamic knowledge graph, multi-modal deep learning and reinforcement learning technology, and the specific advantages are at least as follows:

[0158] 1. By building a unified STKG-Bag and providing transparent decision-making basis, it helps to break down data silos, promote trust and efficient collaboration among airlines, airports and ground services, and fully utilize operational data.

[0159] 2. In the case of data loss or conflict, the DRL-CRQ engine can intelligently deduce the most likely baggage path based on constraints and historical experience, providing key evidence for responsibility attribution and effectively solving the "black box" problem.

[0160] 3. Through M2ARD-Net multi-modal information fusion and hierarchical attention, it can identify the real breakpoint of the responsibility chain from complex correlations and subtle signals, combined with explainability analysis, to improve the accuracy of responsibility determination.

[0161] 4. Using GNN-Pay model can learn and recommend multi-party compensation ratio based on complex responsibility interaction graph, taking into account historical cases, industry rules and specific circumstances of current events, to improve the fairness of compensation and negotiation efficiency.

[0162] 5. The automated and intelligent process shortens the investigation and decision-making time, reduces manual intervention, and accelerates compensation, thereby improving passenger experience and reducing operating costs.

[0163] The above is only the preferred embodiment of the present application, and does not limit the patent scope of the present application, any equivalent structure or equivalent process transformation using the content of the present application specification and drawings, or direct or indirect application in other related technical fields, are also included in the patent protection scope of the present application.

Claims

1. An artificial intelligence-based airport baggage abnormality claim process monitoring method, characterized by, The method comprises the following steps: Real-time acquisition and fusion of baggage transportation data from multiple source heterogeneous systems, through knowledge extraction, entity alignment and relationship inference, construct and dynamically maintain a baggage transportation spatio-temporal knowledge graph STKG-Bag, the STKG-Bag includes entities and spatio-temporal correlation between entities, the entities include baggage, passengers, flights, carriers, operators, geographic locations, time points, operation events and abnormal events; when the baggage abnormal event triggers, extract the spatio-temporal subgraph and operation event sequence related to the abnormal baggage from the STKG-Bag, and input the spatio-temporal subgraph and operation event sequence into a pre-trained multi-modal attention fusion deep network M 2 ARD-Net, the M 2 ARD-Net obtains the responsibility chain breakpoint analysis result by encoding the graph structure context, the time sequence event sequence and the operation specification and abnormal mode prior knowledge, and using multi-modal feature fusion and hierarchical attention mechanism; If the responsibility chain breakpoint analysis result shows that the actual trajectory of the luggage is unclear in a specific area, a dynamic trajectory reconstruction and multi-party responsibility quantification DRL-CRQ process based on deep reinforcement learning is started to obtain a reconstructed path and a compensation scheme; The luggage trajectory in the STKG-Bag, the result of the breakpoint analysis of the responsibility chain, the reconstructed path, and the compensation scheme are interactively displayed through an integrated multi-dimensional visualization interface, and after receiving the final compensation scheme input by a user, the final compensation scheme is used to update the STKG-Bag, M 2 The ARD-Net model, the DRL-CRQ agent, and the GNN-Pay model are subjected to continuous incremental learning and iterative optimization. The process of starting the dynamic trajectory reconstruction and multi-party responsibility quantification DRL-CRQ process based on deep reinforcement learning to obtain the reconstructed path and the compensation scheme specifically comprises: A dynamic environment simulating the luggage circulation is constructed based on the STKG-Bag, operation rules and breakpoint information; An actor-critic framework is adopted by a deep reinforcement learning agent in combination with Monte Carlo tree search auxiliary exploration to learn a strategy in the environment, and N most likely luggage movement paths and their probabilities are reconstructed according to a fine design reward function; For each high-probability reconstructed path, a responsibility-related party interaction graph is constructed, and a graph neural network pay proportion allocation model GNN-Pay is used to quantify the compensation responsibility of each related party to output an adaptive multi-party compensation proportion allocation scheme. The reward function of the deep reinforcement learning agent comprehensively considers path continuity, time consistency, abnormality identification explanation, satisfaction of operation rules and physical constraints, degree of agreement with historical data patterns and path complexity.

2. The method of claim 1, wherein, Knowledge extraction from luggage transportation data includes extracting entities and events from unstructured text by using a Transformer-based pre-trained language model in combination with a conditional random field or a pointer network, and relationship inference includes link prediction by using a graph neural network.

3. The method of claim 1, wherein, The M 2 The spatio-temporal context encoding module STCE-Module in ARD-Net includes a graph structure context encoder GSC of a graph attention network GAT for encoding the topological structure of the baggage transportation subgraph and the dependencies between nodes; a time series event sequence encoder TESE constituted by a bidirectional gated recurrent unit Bi-GRU with a self-attention mechanism for capturing long-term dependencies and time dynamics of the operation event sequence; and a prior knowledge encoder PKE-Module for encoding the operation specification and historical abnormal patterns as prior knowledge and aligning them with the current event features.

4. The method of claim 3, wherein, The M 2 The multi-modal feature fusion and hierarchical attention module M of ARD-Net 2 The FA-Module adopts a gating fusion unit to perform early fusion on the multi-modal features from the STCE-Module, then realizes deep interaction learning between the features through a cross-modal Transformer encoder, and applies a hierarchical fine-grained attention mechanism to locate anomalies from macro sections to micro events.

5. The method of claim 1, wherein, The M 2 The responsibility attribution and breakpoint confidence evaluation module RAD-Module adopts a multi-task learning head to predict the breakpoint position probability, the abnormal reason code and the responsibility subject preliminary inclination score in parallel, and combines an explainability method to highlight the input features that contribute greatly to the decision.

6. The method of claim 1, wherein, The input of the graph neural network pay proportion allocation model GNN-Pay is a responsibility-related party interaction graph constructed based on a high-probability reconstructed path, wherein the node features include the preliminary responsibility inclination score of the related party, the participation degree of the related party in the reconstructed path, the historical operation failure rate and the specification compliance degree, the model outputs the compensation percentage of each related party, and is trained by supervised learning of historical cases and industry rule constraints.

7. The method of claim 1, wherein, The continuous incremental learning and iterative optimization include selecting high-value samples for manual labeling by using an active learning strategy, and evolving and updating the knowledge graph using newly confirmed cases, so as to improve the accuracy of the M 2 ARD-Net, DRL-CRQ agent and GNN-Pay model are periodically or event triggered for incremental training and parameter fine-tuning.

8. An artificial intelligence based airport baggage abnormal claim process monitoring system characterized in that, The computer program is executed by the processor to realize the airport luggage abnormal claim process monitoring method based on artificial intelligence according to any one of claims 1 to 7.

9. A computer-readable storage medium having stored thereon a computer program, characterized in that, The computer program is executed by the processor to realize the airport luggage abnormal claim process monitoring method based on artificial intelligence according to any one of claims 1 to 7.

Images