Cross-role collaborative quantification and instruction system and method for multi-source data fusion

Abstract

The present application relates to the technical field of data processing, in particular to a multi-source data fusion cross-role collaborative quantification and instruction system and method, comprising: a data acquisition module, which acquires multi-source heterogeneous data of multiple operation roles; an individual characteristic representation module, which constructs an individual characteristic vector; a collaborative behavior knowledge base module, which stores a standardized behavior evaluation scale containing observable behavior items and hierarchical scoring criteria; a collaborative state quantification module, which calculates collaborative quantification characteristics based on the individual characteristic vector and the observable behavior items; a feature node identification module, which identifies global extreme value feature nodes, and sorts the contribution degrees of the global extreme value feature nodes and the collaborative quantification characteristics to the abnormal prediction of a target operation system; and an instruction generation module, which retrieves the observable behavior items based on the contribution degree sorting and generates a structured behavior correction instruction set.The present application realizes the quantitative representation and abnormal attribution of the cross-role collaborative state, and establishes an automatic mapping link from quantitative analysis to executable instructions.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a cross-role collaborative quantization and instruction system and method for multi-source data fusion. Background Technology

[0002] In complex systems where multiple roles collaborate, the overall reliability and efficiency depend on the collaborative interaction between data processing subsystems corresponding to various professional positions. However, under the current technological framework, these subsystems operate independently at both the physical storage medium and logical data model levels, forming a heterogeneous data isolation structure. A unified data interaction interface is typically not defined between these subsystems, nor is a standardized data sharing protocol established, resulting in heterogeneity in data encoding formats, timestamp alignment methods, and semantic tagging systems.

[0003] Existing technologies further lack the ability to quantify feature extraction and state characterization of cross-subsystem collaborative interaction processes. Most current data analysis systems designed for multi-role collaborative systems are limited to a single job role or data source. Data inputs only receive operational parameters or status signals from a single role, neglecting the collaborative quality characteristics of interactions between multiple data subsystems. This technological gap results in over 70% of anomalies traceable to cross-subsystem interactions going undetected or unannounced by existing anomaly detection modules before they occur.

[0004] The aforementioned dual technical dilemma of lacking data isolation and collaborative feature extraction methods is specifically reflected in the following two aspects.

[0005] First, regarding the acquisition and fusion processing of multi-source heterogeneous data, existing systems only have single data source access capabilities and lack a processing module for unified acquisition and feature alignment of cross-subsystem interactive data. Existing data processing systems typically design dedicated access channels for a single data source, without reserving interfaces in their architecture for accessing interactive data generated by other subsystems (such as communication messages, handover records, and collaborative calls between different roles). This interactive data is stored in different encoding formats on different physical nodes, and its timestamp base and semantic tag system are heterogeneous. Existing systems lack a unified data pipeline for acquisition, time alignment, semantic alignment, and fusion processing.

[0006] Second, in condition monitoring and anomaly detection systems, the output of anomaly detection results has an inherent lag, and the results cannot be directly mapped to executable instruction parameters for downstream control modules. Current condition monitoring systems primarily rely on event report databases, post-event investigation records, and periodic audit data. All of these data sources have a post-event recording attribute, resulting in a post-event analysis mode for the system's detection. This means the generation of the detection result always lags behind the occurrence of the abnormal event, or lags behind the point when a specific behavioral deviation accumulates to a detectable threshold. Furthermore, even after the detection result is generated, existing systems lack a mapping module to automatically convert it into control instruction parameters for downstream systems. The detection result is typically expressed in the form of high-level semantic tags (such as "communication timing disorder" or "information transmission missing"), and subsequent control decisions depend on the operator's experiential interpretation and manual coding of the detection result to generate control instructions.

[0007] A deeper problem lies in the fact that, due to the lack of a data processing module for quantitative feature extraction of cross-subsystem collaborative states, the input feature dimensions of existing monitoring systems are inherently limited to single-data source behavior deviations and procedure compliance checks. Furthermore, no dedicated feature extraction and monitoring channel is provided for collaborative interface anomalies, a significant source of system-level anomalies. The system's output judgment results (such as "communication response delay sequence deviates from the baseline range" or "high field missing rate in the information transmission link") belong to high-level semantic descriptions. Downstream control operations that can substantially alter the collaborative process require low-level quantifiable command parameters (such as adjusting the communication timing window width or modifying the data transmission verification step size). Currently, the conversion from high-level semantic judgment results to low-level command parameters relies entirely on the empirical analysis and parameter settings of human operators. This conversion process not only introduces time delays but also reduces the repeatability and calibrability of command parameters due to the subjective selection of human parameters, thus making it impossible to verify the downstream control effect through quantitative feedback.

[0008] As a specific application example, the aforementioned technical problems are particularly prominent in aviation operation systems. Aviation operation systems are typical high-reliability, complex systems whose normal operation relies on the collaborative interaction between data processing subsystems corresponding to multiple professional positions, such as pilots, dispatchers, air traffic controllers, maintenance engineers, and flight attendants. Under the existing aviation technology system, pilot operation records and training data are stored in the flight department's data system; dispatcher operation qualification data belongs to an independent database of the operation control system; air traffic controller behavior status parameters are maintained separately by a dedicated air traffic control system; and maintenance engineer technical resume data is incorporated into the maintenance engineering information system. No unified data interaction interface has been defined between these subsystems, nor has a standardized data sharing protocol been established. Existing aviation operation data analysis systems are limited to single-position boundaries. For example, pilot operation feature extraction systems mainly rely on flight parameter time-series data output from flight quality monitoring channels, while air traffic controller behavior analysis systems focus on voice command text streams from air-to-ground communications. Neither can access cross-positional interactive data such as data link communication messages between dispatchers and pilots, handover voice records between maintenance engineers and pilots, and cabin collaborative communications between flight attendants and pilots. Aviation safety data monitoring systems rely on post-incident data sources such as safety incident report databases and after-the-fact investigation records of unsafe incidents. However, the results are often delayed and cannot be automatically mapped to training instructions or parameter adjustments. This example clearly illustrates the real-world existence of the aforementioned common technical problems.

[0009] In summary, the core defects of the existing technology system can be summarized as follows: the data subsystems corresponding to multiple roles are isolated at the physical and logical levels, and there is a lack of data processing modules for quantitative feature extraction of cross-subsystem collaborative interaction processes. This causes the entire status monitoring chain to remain at the level of anomaly detection and report generation, and it is impossible to achieve the technological leap from the lagging feature extraction of events that have occurred to the situation prediction of events that have not occurred, and from single data source behavior analysis to multi-data source collaborative status analysis. Summary of the Invention

[0010] To address the aforementioned technical problems, the technical solution adopted by this invention is as follows:

[0011] According to a first aspect of the present invention, a cross-role collaborative quantization and instruction system for multi-source data fusion is provided, the system comprising:

[0012] The data acquisition module is configured to collect multi-source heterogeneous data from multiple operating roles through a dedicated data interface. The multi-source heterogeneous data includes at least runtime data, training data, and interaction data between operating roles.

[0013] The individual feature representation module is configured to extract multi-dimensional operational capability features of each operational role from the multi-source heterogeneous data based on a preset feature analysis framework, and construct an individual feature vector.

[0014] The collaborative behavior knowledge base module is configured to store standardized behavior assessment scales for collaborative scenarios involving multiple operational roles. The standardized behavior assessment scales include multiple observable behavior items and their corresponding graded scoring criteria.

[0015] The collaborative state quantization module is configured to receive the individual feature vector output by the individual feature representation module, and call observable behavior items that match the current scene from the collaborative behavior knowledge base module, and calculate the collaborative quantization features between at least one pair of operation roles based on the two.

[0016] The feature node identification module is configured to construct a multi-index comparison model across operational roles and dimensions based on the individual feature vectors, identify global extreme value feature nodes, and input the global extreme value feature nodes and the collaborative quantization features into an interpretable machine learning model to output the contribution ranking of each feature to the anomaly prediction of the target operating system.

[0017] The instruction generation module is configured to determine the target feature dimension based on the contribution ranking, retrieve observable behavior items that have a mapping relationship with the target feature dimension from the collaborative behavior knowledge base module, and automatically generate a structured behavior correction instruction set based on the retrieved observable behavior items and their hierarchical scoring criteria.

[0018] According to a second aspect of the present invention, a cross-role collaborative quantization and instruction method for multi-source data fusion is provided, comprising the following steps:

[0019] Multi-source heterogeneous data from multiple operational roles is collected through a dedicated data interface. The multi-source heterogeneous data includes at least runtime data, training data, and interaction data between operational roles.

[0020] Based on a preset feature analysis framework, multi-dimensional operational capability features of each operational role are extracted from the multi-source heterogeneous data to construct individual feature vectors.

[0021] A pre-built collaborative behavior knowledge base is invoked, which stores a standardized behavior assessment scale containing multiple observable behavior items and their rating criteria; based on the individual feature vectors and observable behavior items from the collaborative behavior knowledge base that match the current scenario, at least one pair of operational roles is used to calculate the collaborative quantitative features.

[0022] Based on the individual feature vectors, a multi-indicator comparison model is constructed across operational roles and dimensions to identify global extreme value feature nodes.

[0023] The global extreme value feature node and the collaborative quantization feature are used as input features and fed into an interpretable machine learning model. The model outputs a ranking of the contributions of each feature to the anomaly prediction of the target operating system.

[0024] Based on the contribution ranking, the target feature dimension is determined. Observable behavior items that have a mapping relationship with the target feature dimension are retrieved from the collaborative behavior knowledge base. Based on the retrieved observable behavior items and their hierarchical scoring criteria, a structured behavior correction instruction set is automatically generated.

[0025] This application solves the data silo problem caused by heterogeneous data formats and missing interfaces in existing subsystems by uniformly collecting operational data, training data, and interaction data from multiple operational roles through a data acquisition module, providing a complete data foundation for cross-role collaborative analysis. The collaborative state quantification module calculates the collaborative quantification features between operational role pairs based on individual feature vectors and observable behavior items, achieving for the first time a quantitative characterization of cross-role interaction quality, overcoming the limitation of existing systems that can only analyze single-role operational parameters. The feature node identification module inputs global extreme value feature nodes and collaborative quantification features into an interpretable machine learning model, outputting a ranking of the contribution of each feature to the anomaly prediction of the target operating system, achieving quantifiable and interpretable anomaly attribution. The instruction generation module retrieves corresponding observable behavior items based on contribution ranking and automatically generates a structured behavior correction instruction set, establishing an automated mapping link from anomaly detection to executable instructions, eliminating the time delay and subjective bias caused by manual conversion in existing technologies.

[0026] It should be understood that the description in this section is not intended to identify key or essential features of the embodiments of the present invention, nor is it intended to limit the scope of the invention. Other features of the invention will become readily apparent from the following description. Attached Figure Description

[0027] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0028] Figure 1 This is a structural block diagram of a cross-role collaborative quantization and instruction system for multi-source data fusion provided in an embodiment of the present invention;

[0029] Figure 2 A flowchart of a cross-role collaborative quantization and instruction method for multi-source data fusion provided in an embodiment of the present invention. Detailed Implementation

[0030] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0031] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein in the description of this invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0032] It should be noted that some exemplary embodiments are described as processes or methods depicted as flowcharts. Although the flowcharts describe the steps as sequential processes, many of these steps can be performed in parallel, concurrently, or simultaneously. Furthermore, the order of the steps can be rearranged. A process can be terminated when its operation is complete, but it may also have additional steps not included in the figures. A process can correspond to a method, function, procedure, subroutine, subroutine, etc.

[0033] Example 1

[0034] This invention provides a cross-role collaborative quantification and instruction system based on multi-source data fusion. Its overall objective is to: collect heterogeneous data from multiple sources, construct individual operational feature representation models and multi-role collaborative state quantification models covering various operational roles, quantify cross-role cognitive alignment and interaction coupling coefficients, identify key abnormal feature nodes in the target operating system using an interpretable machine learning module, and automatically generate hierarchical control suggestions and structured behavioral instruction sequences. This establishes a technical link from feature quantification analysis to instruction parameter mapping, thereby achieving a technological leap from delayed anomaly detection to prior situation prediction, and from single-source data source behavior analysis to multi-source data source collaborative state analysis. Figure 1 As shown, the multi-source data fusion cross-role collaborative quantification and instruction system provided in this embodiment includes: a data acquisition module 1, an individual feature representation module 2, a collaborative behavior knowledge base module 3, a collaborative state quantification module 4, a feature node recognition module 5, and an instruction generation module 6. The structure and function of each module are described in detail below.

[0035] I. Data Acquisition Module

[0036] The data acquisition module 1 is configured to collect multi-source heterogeneous data from multiple operating roles through a dedicated data interface. The multi-source heterogeneous data includes at least running data, training data, and interaction data between operating roles.

[0037] The operational data includes at least one of the following: status monitoring parameters, task timing records, and resource configuration records related to the execution of the operation task. Status monitoring parameters refer to quantifiable indicators characterizing the physical or logical state of the operation object or environment during task execution, including but not limited to motion parameters, position parameters, environmental parameters, and equipment operating status parameters. Task timing records refer to a set of data recording the task execution steps, event trigger times, and duration of each step in chronological order. Resource configuration records refer to the recorded data related to task execution, including operator allocation, equipment allocation, time window allocation, and task priority settings.

[0038] The training data includes at least one of the following: operational behavior records in a simulated environment and competency assessment results. The simulated environment refers to a simulation system or virtual environment used to reproduce real-world operational scenarios or typical task situations, capable of recording operator input and system response data. Operational behavior records refer to operator input commands, operation sequences, operation paths, and responses to changes in the virtual device's state captured in the simulated environment. Competency assessment results refer to the quantitative evaluation of the operator's performance in the simulated environment or real-world scenario based on preset assessment standards.

[0039] The interactive data between operational roles includes at least one of the following: voice communication records and corresponding text transcription, structured data link messages, internal call audio records, and electronic task handover records. Voice communication records refer to real-time voice signals transmitted between operational roles via communication channels and their text transcription after speech recognition. Structured data link messages refer to message data encoded according to a specific data format and communication protocol, including a data payload containing sender identifier, receiver identifier, message type, timestamp, and predefined fields. Internal call audio records refer to audio data of conversations between different seats or terminals within the same operating platform or system. Electronic task handover records refer to structured electronic records that transmit task status, completed items, pending items, and precautions between different operational roles or different operational stages.

[0040] The following example, using the aviation operations field, illustrates the specific sources of the aforementioned data, but the scope of application of this system is not limited to this.

[0041] In a specific application scenario, the system is deployed in an aviation operating environment. At this time:

[0042] (1) Data interface and operation role

[0043] The dedicated data interface is a dedicated data interface for aviation operations, supporting real-time or periodic data collection. The multiple operational roles include at least pilots, dispatchers, air traffic controllers, maintenance engineers, and flight attendants.

[0044] (2) Specific sources of operational data

[0045] Condition monitoring parameters: derived from the flight data recorder, including at least barometric altitude, airspeed, and engine parameters;

[0046] Task timing records: derived from the segment execution timestamps in the flight operation control system;

[0047] Resource configuration records: These are derived from the crew member and aircraft number binding records in the scheduling system.

[0048] (3) Specific sources of training data

[0049] Simulation environment: corresponding to a full-motion flight simulator or a fixed-program trainer;

[0050] Operation behavior records: corresponding to joystick displacement, switch action sequences, etc. on the simulator;

[0051] Competency assessment results: corresponding to the scores in the simulator training evaluation form.

[0052] (4) Specific sources of data for interaction between the operating roles

[0053] The aforementioned data exchanged between operational roles originates from a dedicated aviation operations communication link, specifically including:

[0054] Voice communication recording and text transcription: corresponding to VHF recordings of air-to-ground communication and their transcribed texts generated by speech recognition;

[0055] Structured data link messages: These correspond to the send and receive messages in systems such as ACARS or SATCOM;

[0056] Intercom audio recordings: Audio recordings corresponding to the cockpit voice recorder and the cabin intercom system;

[0057] Electronic task handover record: corresponds to the handover signature information in the electronic work card system and release record.

[0058] The aforementioned data types are connected to the data acquisition module through corresponding dedicated data interfaces. This module performs the following preprocessing operations on the collected multi-source heterogeneous data: unified timestamp alignment (converting the time base of data from different sources into a unified system timeline), data cleaning (removing outliers and filling in missing fields), and structured storage (organizing the data into a data format that can be directly called by downstream modules according to a preset data model).

[0059] Those skilled in the art should understand that the specific data types, sources, and preprocessing methods described above in the aviation operating environment are for illustrative purposes only, intended to fully disclose the implementation of the present invention, and not to limit the scope of application of this system. Provided the technical solution can be generalized and promoted, the system can also be deployed in other high-reliability systems with multiple roles operating collaboratively.

[0060] II. Individual Characteristic Representation Module

[0061] The individual feature representation module 2 is configured to extract multi-dimensional operational capability features of each operational role from the multi-source heterogeneous data based on a preset feature analysis framework, and construct an individual feature vector.

[0062] In this invention, a feature analysis framework is used to define a set of quantitative indicators to characterize the operational capabilities of an operator during task execution. Each dimension corresponds to a quantifiable feature component, and the dimensions are independent of each other or have a pre-defined correlation structure. The feature analysis framework can be obtained based on the decomposition and modeling of key operational behaviors in typical task scenarios.

[0063] Specifically, the individual feature representation module 2 extracts corresponding quantitative features for each operational role from the operational data, training data, and interaction data between operational roles output by the data acquisition module 1, according to the dimensions defined in the feature analysis framework. The extracted quantitative features include, but are not limited to:

[0064] (1) Operation accuracy

[0065] Definition: Operational accuracy characterizes the degree of deviation between the actual operation value and the preset target value of an operator when performing a task, reflecting the ability to follow predetermined parameters or trajectories.

[0066] Acquisition method: Extracted from status monitoring parameters. For tasks requiring the maintenance of a specific parameter range, the deviation metric is calculated based on the difference between the actual parameter trajectory and the standard parameter curve; for tasks requiring discrete operations, the deviation is calculated based on the time deviation between the operation execution time and the specified time, or the spatial deviation between the operation position and the target position. The larger the deviation, the lower the precision quantification value. The deviation value is normalized and mapped to a preset scoring interval to obtain the operation precision quantification value.

[0067] (2) Reaction time

[0068] Definition: Reaction timeliness represents the time delay between the occurrence of a triggering event and the first response by the operating role.

[0069] Acquisition method: Extracted from status monitoring parameters and task timing records. Identify the timestamp of the triggering event (such as alarm signal, command issuance) and the timestamp of the operator's first response action, and calculate the difference between the two. Divide this difference by the maximum allowable delay for this task type or a reference range based on historical data to obtain a normalized reaction time quantification value (the smaller the delay, the higher the score).

[0070] (3) Task switching efficiency

[0071] Definition: Task switching efficiency represents the time cost required for an operator to switch between two different tasks, as well as the transition time required to return to a steady-state operating level after the switch.

[0072] Acquisition method: Extract the end timestamp of the previous task and the start timestamp of the next task from the task timing record, and calculate the difference as the switching time cost. Extract the time required for the operation parameters after switching to enter the steady-state allowable error range from the status monitoring parameters, as the transition time. Normalize the above two timestamps respectively, and then aggregate them according to preset weights to obtain the quantitative value of task switching efficiency (the shorter the time, the higher the efficiency).

[0073] (4) Characteristics of time resource allocation

[0074] Definition: Time resource allocation characteristics represent the consistency between the distribution of time investment ratios of operational roles in various types of tasks and the preset baseline distribution, as well as the frequency of task switching and the proportion of time spent on critical tasks per unit of time.

[0075] Acquisition Method: Extracted from task timeline records and resource configuration records. The task timeline is segmented by task type, and the total duration of each task type is calculated, along with its proportion of the total time, to obtain the actual distribution vector. A preset baseline distribution vector (e.g., based on the average input of efficient operators or expert calibration values) is retrieved from the knowledge base or historical data. The similarity between the actual distribution and the baseline distribution is calculated (using cosine similarity, inverse KL divergence, or Euclidean distance normalization) to obtain a similarity quantification value. Simultaneously, the number of task type changes per unit time is calculated as the task switching frequency quantification value, and the cumulative duration of predefined key task types as a proportion of the total time is calculated as the key task percentage quantification value.

[0076] Determining the final quantified value: The similarity quantified value, the task switching frequency quantified value (after inverse normalization so that higher frequency results in lower scores), and the key task proportion quantified value are weighted and aggregated according to preset weights to obtain the final quantified value of the time resource allocation feature. When it is not necessary to distinguish the above sub-dimensions, the similarity quantified value can be directly used as the quantified value of the time resource allocation feature.

[0077] (5) Operational proficiency

[0078] Definition: Operational proficiency represents the degree to which an operator is familiar with routine operation sequences, reflecting the fluency and accuracy of the operation.

[0079] Acquisition method: Extracted from operational behavior records and competency assessment results in the training data. Specific indicators include: the total time spent completing a specified sequence of operations in a simulated environment, the number of erroneous operations (such as omissions, incorrect sequences, and redundant operations), and the standardized score in the assessment scoring table. The above indicators are weighted and synthesized according to preset weights and normalized to a unified quantification range to obtain a quantified value for operational proficiency.

[0080] (6) Ability to handle abnormal situations

[0081] Definition: Anomaly handling capability characterizes the efficiency and quality of response of an operational role when facing abnormal or emergency scenarios, including at least anomaly detection delay, completeness of handling steps, and recovery time.

[0082] Acquisition method: Extracted from operation response records for abnormal or emergency scenarios in the training data. Anomaly detection latency refers to the time difference between the injection of an anomaly scenario and the first identification and targeted response by the operating role; completeness of handling steps refers to the proportion of the number of handling steps actually executed to the number of steps specified in the standard handling procedure; recovery time refers to the time elapsed from the occurrence of the anomaly to the system recovering to a safe and controllable state. The above sub-dimensions are weighted and aggregated to obtain a quantitative value of anomaly handling capability.

[0083] (7) Integrity of information transmission

[0084] Definition: Information transmission integrity characterizes the completeness of the information content transmitted when an operator sends or receives information during an interaction.

[0085] Acquisition Method: Extracted from interactive data. For structured messages, based on a preset set of required fields, the proportion of missing fields to the total number of required fields is calculated, with completeness = 1 - missing rate. For unstructured text or speech-to-text, based on keyword detection or a semantic completeness model, it is determined whether key information points are fully conveyed, and the proportion of fully conveyed information points to the total number of information points that should be conveyed is calculated. The above completeness is normalized to obtain a quantitative value for information transmission completeness.

[0086] (8) Timeliness of information transmission

[0087] Definition: Information delivery timeliness characterizes the timeliness of information sent by the operating role during the interaction process, and reflects whether the information delivery meets the time requirements of the task.

[0088] Acquisition method: Extracted from the timestamp field of the interaction data. For each information interaction requiring a reply or confirmation, calculate the time interval from the moment the message is sent to the moment the confirmation or reply is received. Compare this interval with a preset reference threshold corresponding to this interaction type (such as the maximum allowable delay or optimal response window based on historical data), and score it according to the degree to which it exceeds the threshold (the less it exceeds, the higher the score). Take the average or weighted average of the scores of all interactions to obtain a quantitative value for the timeliness of information delivery.

[0089] (9) Closed-loop confirmation characteristics

[0090] Definition: Closed-loop confirmation feature indicates whether the operating role performs closed-loop confirmation of the received information during the interaction process, reflecting the reliability of communication and the sense of responsibility.

[0091] Acquisition Method: After each message sending event, detect whether a corresponding receipt confirmation or feedback event exists in the interaction data. Confirmation events can be explicit receipts, confirmation phrases (such as "received" or "understood"), or status feedback after task execution. For each sending event, if a valid confirmation event is detected within a preset time window, it is counted as a closed loop; otherwise, it is considered an open loop. Calculate the closed-loop confirmation rate = number of closed-loop events / total number of sending events; this ratio is directly used as the quantitative value of the closed-loop confirmation characteristic.

[0092] After normalization, the aforementioned quantitative features are arranged in dimensional order as defined by the feature analysis framework to form an individual feature vector. This vector uniquely represents the distribution of operational capabilities for the current role within the current evaluation period. The calculation of each feature can be standardized based on a preset benchmark value or historical data distribution to ensure the comparability of features across different operational roles and task scenarios.

[0093] The individual feature representation module continuously receives newly collected data. Whenever new running records, training records, or interaction data are accessed, it automatically recalculates the quantitative indicators of the affected dimensions and refreshes the corresponding individual feature vectors, thereby achieving a gradual update of the individual feature representation.

[0094] In a preferred example (not limiting), the feature analysis framework may adopt a capability-based training and assessment framework recognized by the International Civil Aviation Organization (ICAO) or an equivalent industry standard. In this example, operational capability dimensions can be divided into: technical operational capabilities (such as flight control precision and procedural execution standardization), non-technical operational capabilities (such as situational awareness, decision-making, and communication and coordination), and job adaptability characteristics (such as workload management and task prioritization). The above division is for illustrative purposes only; the feature analysis framework of this invention is not limited to this and other dimensions can be customized or extended according to the actual system deployment scenario.

[0095] Through the above configuration, the individual feature representation module 2 transforms the original multi-source heterogeneous data into structured individual feature vectors, providing a standardized input data format for subsequent collaborative state quantification and feature node identification.

[0096] III. Collaborative Behavior Knowledge Base Module

[0097] The collaborative behavior knowledge base module 3 is configured to store standardized behavior assessment scales for collaborative scenarios involving multiple operational roles. These standardized behavior assessment scales include multiple observable behavior items and their corresponding grading criteria.

[0098] In this embodiment, the standardized behavioral assessment scale adopts a three-layer structure:

[0099] First layer: Collaborative behavior dimension layer

[0100] Define multiple collaborative behavior dimensions common to different operational roles, including at least:

[0101] Information synchronization dimension: This refers to the proactive sharing of key information related to the current task, such as status, plans, and environmental changes, among operational roles to ensure that all parties have a consistent understanding of the task situation. This dimension reflects the timeliness, accuracy, and proactivity of information sharing.

[0102] Clarification of Intent Dimension: This refers to the action of an operator, after issuing instructions, making suggestions, or expressing judgments, to proactively confirm whether the other party has correctly understood their intent, or to proactively request further explanation when receiving ambiguous information. This dimension reflects the clarity of communication and the two-way confirmation mechanism.

[0103] Closed-loop confirmation dimension: This refers to the actions taken by the operating role after receiving instructions or key information, including providing feedback on the receipt result or fulfilling commitments to the sender, and informing relevant parties after task completion or status change. This dimension reflects the closure of the communication link and the completeness of responsibility transfer.

[0104] Proactive Support Dimension: This refers to the actions taken by an operator to proactively provide resources, information, or operational assistance to other operators based on their assessment of the overall task situation, without receiving an explicit request. This dimension reflects the operator's forward-thinking approach and team orientation.

[0105] Conflict resolution dimension: This refers to the behavior of resolving conflicts and reaching consensus through negotiation, compromise, or escalation of decisions when operational roles experience differences in goals, resource competition, or disagreements in judgment. This dimension reflects the effectiveness of conflict resolution and collaborative resilience.

[0106] The above dimensions are extracted based on the interaction patterns of collaborative tasks involving multiple operational roles. They do not depend on specific combinations of operational roles or operating scenarios and have universality across positions and scenarios.

[0107] Second layer: Observable behavior item layer

[0108] Under each collaborative behavior dimension, multiple observable behavior items corresponding to that dimension are stored. Each observable behavior item is configured with:

[0109] Behavioral description text: A standardized description of the specific actions or verbal performance of the behavior;

[0110] Source document basis: the corresponding clause index in the standard operating procedures, training manuals or industry standards that support the definition of this behavior item;

[0111] Grading Criteria: A behavior-anchored grading method is used, defining multiple performance levels for each observable behavior item. These levels are arranged sequentially from the lowest level (i.e., behavior absent or not exhibited) to the highest level (i.e., behavior excellent or exemplary). Each level is configured with:

[0112] Observable feature description: The specific behavioral characteristics used to determine this level of behavior;

[0113] Scoring range: The range of quantitative scores corresponding to this level.

[0114] With the above configuration, the scoring process uses observable features as an objective reference, reducing subjective bias in scoring.

[0115] Third layer: Context adaptation layer

[0116] Storage context adaptation mapping rules are used to define the correspondence between each observable behavior item and a specific operation role in a specific operating scenario.

[0117] At the general level, the specific operational roles cover combinations of two types of functional roles: task issuer and task executor, information provider and information user, and resource coordinator and resource requester. The specific operational scenarios include, but are not limited to: collaborative handling under abnormal conditions, multi-party negotiation under resource-constrained conditions, multi-stage task handover, and rapid decision-making in emergency situations.

[0118] The following example uses the aviation operating environment (this does not constitute a limitation on the scope of application of the system):

[0119] Examples of specific operational role pairs: combinations of the first operational role and the second operational role, specifically corresponding to pilot and dispatcher, pilot and controller, pilot and maintenance engineer, pilot and flight attendant, dispatcher and controller.

[0120] Examples of specific operational scenarios: collaborative decision-making for rerouting when weather conditions deviate from expectations, emergency consultation when fuel status is close to the warning line, collaborative de-icing operations under ground icing conditions, multi-position release handover under equipment failure retention conditions, and selection of alternate airports when the destination airport is unavailable.

[0121] This three-layer structure decouples the evaluation logic from the scenario configuration: when the system needs to be expanded to new combinations of operational roles or new operating scenarios, only the corresponding mapping rules need to be added to the third layer, without modifying the dimension definitions in the first layer or the observable behavior item library and scoring criteria in the second layer. This architecture improves the system's scalability and reusability.

[0122] IV. Cooperative Status Quantification Module

[0123] The collaborative state quantization module 4 is configured to receive the individual feature vector output by the individual feature representation module, and call the observable behavior items that match the current scene from the collaborative behavior knowledge base module, and calculate the collaborative quantization features between at least one pair of operation roles based on the two.

[0124] In this invention, collaborative quantification features are used to quantitatively characterize the information exchange quality, cognitive consistency, and task coordination efficiency exhibited by two or more operational roles during collaborative interaction. The collaborative quantification features include at least: the state alignment rate for a three-party operational role combination, and the collaborative coupling coefficient for a two-party operational role pair.

[0125] For operational role combinations involving three parties—a task coordinator, an instruction issuer, and an operation executor—the collaborative state quantification module 4 employs a Hidden Markov Model to jointly model the interaction sequence of the three parties, calculating the state alignment rate as a shared cognitive metric. The specific calculation process is as follows:

[0126] Step 1: Observation Sequence Encoding

[0127] The interaction behaviors of the three parties on the timeline (including communication event types, operation response delays, task status changes, etc.) are discretized according to fixed time windows and encoded into observable symbol sequences, which are used as the observation inputs for the three parties respectively.

[0128] Step 2: Hidden State Decoding

[0129] The above observation sequence is input into a pre-trained Hidden Markov Model (HMM). The structure of the HMM is as follows: each role corresponds to a hidden state variable, and the joint values ​​of the three hidden states constitute a combined state space (for example, if each role has K hidden states, the number of joint states is K). 3The Hidden Markov Model (HMM) uses the Viterbi algorithm to decode the joint hidden state at each time step, obtaining the hidden state sequences of the three parties at each time step t, denoted as follows: Where t ranges from 1 to T, and T is the total number of time steps. The semantic definitions of each hidden state are as follows (taking a three-level division as an example; the specific number of levels can be adjusted according to the actual scenario):

[0130] s=1: Low collaborative cognition (low cognitive consistency among roles, with interaction behaviors exhibiting more confirmation delays or missing information).

[0131] s=2: Partial collaborative cognition (there is a certain degree of consistency in cognition among roles, but they have not yet reached complete synchronization).

[0132] s=3: Fully collaborative cognition (highly consistent cognition among roles, smooth interaction and timely feedback).

[0133] The semantic definitions above are for illustrative purposes only and do not constitute a limitation on the scope of protection of this system.

[0134] Step 3: State Alignment Calculation

[0135] State alignment rate (SMM) is defined as the proportion of time steps in which the three hidden states are completely consistent at the same time step out of the total number of time steps: .

[0136] Where I[·] is an indicator function, which takes the value 1 when the condition inside the parentheses is true, and 0 otherwise.

[0137] Step 4: Collaborative Early Warning Judgment

[0138] When the state alignment rate is lower than the dynamic adjustment threshold, the cooperative state quantization module outputs a cooperative early warning signal. The dynamic adjustment threshold is automatically calculated based on the percentile (or quartile) of the state alignment rate distribution in historical normal cooperative data, and can be set according to the operational phase (such as takeoff, cruise, and approach).

[0139] In this embodiment, the training method for the Hidden Markov Model can employ existing methods, such as:

[0140] The three-party interaction behavior sequences in historical operation data are collected, and the Baum-Welch algorithm (an expectation-maximization algorithm) is used to estimate parameters on unlabeled data. The goal is to maximize the generation probability of the observation sequence and estimate the model parameters, including: the initial state probability vector, the state transition probability matrix, and the observation symbol emission probability matrix.

[0141] Optionally, if the historical data includes annotations of collaborative states (such as based on investigation conclusions of insecurity incidents or expert ratings), the model parameters can be fine-tuned using supervised or semi-supervised methods. After training, the model parameters are fixed and used for real-time decoding.

[0142] For operation role pairs that consist of only two parties, this module calculates the collaboration coupling coefficient based on the following four sub-indicators:

[0143] Interaction frequency: The number of communications or message exchanges between two parties per unit of time;

[0144] Confirmation rate: In interactions requiring confirmation, the percentage of times the receiver returns confirmation to the sender;

[0145] Response latency: The time interval between sending a message and receiving the first response;

[0146] Completeness of handover rate: The percentage of mandatory fields and essential steps that are fully executed during the task handover process.

[0147] The formula for calculating the cooperative coupling coefficient CCC is: CCC = w f •F norm +w c •c+w τ •exp(-λ·τ)+w h ·h.

[0148] Among them, w f w c w τ w h These are weighting coefficients, satisfying a sum of 1 and all being non-negative. Default values ​​can be preset based on domain experience and can be dynamically adjusted. In one specific implementation, the weighting coefficients adopt the following default values: communication frequency weight w f =0.3, information confirmation rate weight w c =0.2, response delay weight w τ =0.25, task handover integrity rate weight w h =0.2. These default values ​​are set based on expert questionnaire survey results on the impact of various indicators on the quality of collaboration in typical operating scenarios. In actual use, the system supports automatically switching different weight preset values ​​according to the current mission stage (such as takeoff, cruise, approach) or operating scenario type (such as normal situation, abnormal procedure), allowing for dynamic adjustment.

[0149] F norm Let f be the normalized value of the interaction frequency, and let f be the actual number of interactions per unit time. optTo determine the optimal number of interactions for the current task type (i.e., the reference frequency that is most conducive to collaborative efficiency in this type of task scenario, based on historical data or expert experience), the normalized frequency score F is calculated. norm Calculate F using the following formula: norm = , where σ f The parameter controlling the Gaussian kernel width is used to adjust the sensitivity to frequency deviation. This Gaussian normalization function maps the actual frequency f to the interval [0,1]. When f=f opt At that time, F norm =1 (highest score); when f deviates from f opt At that time, F norm The score decreases as the frequency of communication decreases along a Gaussian curve, with further deviations resulting in lower scores. This reflects the technical concept that "too low a frequency may indicate insufficient communication, while too high a frequency may indicate redundant interference."

[0150] c represents the confirmation rate, ranging from [0,1]; τ represents the response latency; λ represents the latency sensitivity coefficient (calibrated based on historical data), which uses a negative exponential function to make the contribution greater as the latency decreases; h represents the handover integrity rate, ranging from [0,1]. The CCC value ranges from [0,1], and a larger value indicates higher quality information exchange and smoother task connection between the operation roles.

[0151] In a further configuration of this module, the cooperative coupling coefficient CCC serves two levels:

[0152] Direct quantification: Serves as a direct quantitative indicator of the quality of collaboration between any two operational roles, and is used for subsequent collaboration risk assessment.

[0153] Networked modeling: The edge weights, used as initial values ​​for constructing the graph neural network model, are output as a global collaborative efficiency coefficient after graph convolution and global pooling. This coefficient reflects the overall efficiency level of the multi-operational collaborative network and is used to adjust the state transition probability matrix parameters in the aforementioned hidden Markov model.

[0154] This forms a three-level progressive evaluation architecture from "pair coupling coefficient → global collaborative efficiency coefficient → state alignment rate", enabling the system to quantitatively characterize the collaborative state of multiple operational roles from three levels: micro (pair interaction quality), meso (global network efficiency), and macro (cognitive consistency).

[0155] Furthermore, the collaborative state quantization module 4 is also configured to: construct a graph neural network (GCN) model with each operation role as a node and the collaborative coupling coefficient as the edge weight, update node features through graph convolution operations, and output the global collaborative efficiency coefficient after global pooling.

[0156] As a specific implementation, a Graph Convolutional Network (GCN) is used. The basic structure of GCN (including graph convolutional layers, neighbor aggregation, symmetric normalization, etc.) is well-known in the art, and its general principles will not be detailed here. The following focuses on the specific configuration of GCN in this invention:

[0157] Let N be the number of roles involved in the collaboration. Each node v i The initial eigenvector x i ∈R dv It is composed of the individual feature vector of the role output by the individual feature representation module. Node v i With v j The edge weight e between ij Let CCC be the collaboration coupling coefficient for the corresponding operation role pair. The values ​​of i and j are both from 1 to N.

[0158] That is, the hidden state vector of all nodes in the r-th layer is h. (r) i , where r = 0, 1, ..., L-1, and L is the total number of layers in the graph convolutional network (L = 2 in this embodiment). Initial state h (0) i =x i The (r+1)th layer node v i The hidden state vector is updated as follows: .

[0159] Where: N(v) i ) is node v i The set of neighboring nodes; d i For node v i The weighted degree with self-loop, ;d j For node v j The weighted degree with self-loop, N(v) j ) is node v j The set of neighboring nodes, e ju Let w be the edge weight between node j and its neighbor u. r Let r be the learnable weight matrix of the r-th layer. h is a non-linear activation function. 0 i =x i These are the initial node features.

[0160] After L layers of graph convolution, the final embedding vector h of all nodes is... (L) i Weighted average pooling is performed based on the node's weighted degree to obtain the global embedding vector z:

[0161] .

[0162] in For node v i The neighbor weighted degree.

[0163] The global embedding vector z is input into the fully connected layer, and the global co-efficiency coefficient α is output through the Sigmoid activation function: α = Sigmoid(w o T ·z+b o ).

[0164] Where w o and b o These are the weight vector and bias term of the output layer, respectively, α∈[0,1]. This coefficient comprehensively reflects the overall collaborative efficiency level of the multi-role collaborative network composed of all current operational roles.

[0165] The graph convolutional network described above is trained in a supervised manner using historical operational data. Training labels can be based on a tiered classification of insecurity events (e.g., classifying events into three or five levels of severity) or expert ratings of collaborative efficiency. Training employs a backpropagation algorithm, with the optimization objective being to minimize the error between the predicted α and the true label. After training, the model parameters are fixed for real-time inference.

[0166] In this embodiment, the global collaborative efficiency coefficient α is fed back into the aforementioned hidden Markov model used for measuring the collaborative performance of the three parties, and is used to adjust the state transition probability matrix of the model. In an illustrative embodiment, the specific adjustment method is as follows:

[0167] The initial value of the state transition probability matrix A of the Hidden Markov Model can be estimated in advance using historical data (such as statistics under normal cooperative states).

[0168] During real-time inference, the transition probability matrix is ​​dynamically scaled based on the currently calculated α value:

[0169] For the probability of transitioning from the current state to a higher level of collaborative cognition (e.g., from s=1 or s=2 to s=3), multiplying by α (or an increasing function of α) makes the system more inclined to predict transitions to higher collaborative states when the global collaborative efficiency is higher.

[0170] The probability of transitioning to a lower level of collaborative cognition is multiplied by 1-α or its decreasing function.

[0171] Another implementation is to use α as a prior parameter in the Hidden Markov Model (HMM) to directly adjust the posterior distribution of the state transition probability matrix, i.e., Ac = Normalize(A⊙M(α)), where Ac represents the normalized state transition probability matrix adjusted by α, i.e., the updated transition matrix incorporating current global collaborative efficiency information. During real-time decoding, the HMM will use Ac instead of the original A, allowing the state decoding results to dynamically respond to changes in the overall efficiency of the multi-role collaborative network. ⊙ represents element-wise multiplication, and M(α) is the scaling matrix generated by α. Normalize represents row (or column) normalization of the matrix so that the sum of the transition probabilities in each row is 1.

[0172] In one illustrative embodiment, M(α) is generated as follows:

[0173] Assume the Hidden Markov Model has K hidden states (e.g., K=3), with state numbers representing increasing levels of collaborative cognition. Define the elements of the adjustment matrix M(α) as follows: for any two hidden state numbers p and q, if q is greater than p (i.e., transitioning from state p to state q represents an increase in cognitive level), the corresponding element is α plus a small smoothing constant β; if q is less than p (i.e., transitioning from state p to state q represents a decrease in cognitive level), the corresponding element is 1 minus α plus another small smoothing constant γ; if p equals q (i.e., maintaining the current state), the corresponding element is 1. Here, β and γ are preset positive decimals used to ensure all elements are greater than zero.

[0174] Another simpler implementation: directly construct the diagonal matrix M(α) = diag(1, α, α) 2 or similar scaling vectors, but usually require appropriate normalization.

[0175] Through the above adjustments, the state decoding result of the Hidden Markov Model no longer depends solely on the historical statistical transition patterns, but can integrate the current global collaborative efficiency information in real time, making the state alignment rate (shared cognitive metric) more sensitive to the current collaborative situation.

[0176] V. Feature Node Recognition Module

[0177] The feature node identification module 5 is connected to the individual feature representation module 2 and the collaborative state quantization module 4 respectively. It is configured to: construct a multi-index comparison model across operation roles and dimensions based on the individual feature vector, identify global extreme value feature nodes, and input the global extreme value feature nodes and the collaborative quantization features into an interpretable machine learning model, and output the contribution ranking of each feature to the anomaly prediction of the target operating system.

[0178] Specifically, this module includes the following processing flow executed sequentially:

[0179] (I) Identification of global extreme value feature nodes (extreme value location of individual operational capability feature dimension)

[0180] The feature node recognition module 5 receives the individual feature vectors of all operational roles output by the individual feature representation module 2. Each individual feature vector contains multiple operational capability feature dimensions, such as: operational accuracy, reaction timeliness, task switching efficiency, time resource allocation features, operational proficiency, anomaly handling ability, information transmission integrity, information transmission timeliness, and closed-loop confirmation features.

[0181] To eliminate the problem of incomparability between different feature dimensions due to differences in units and value ranges, the feature node recognition module first normalizes the original quantized values ​​of each feature dimension, mapping all values ​​to the same numerical space (e.g., the [0,1] interval). The normalization method can adopt Min-Max standardization, where the minimum and maximum values ​​of each dimension are pre-set based on historical data distribution or domain knowledge.

[0182] Building upon this foundation, this module constructs a multi-indicator comparison model. This model treats the normalized quantified values ​​of all feature dimensions for all operational roles as a flat list of values, each with two labels: operational role identifier (e.g., role number or role type) and feature dimension name (e.g., "closed-loop confirmation feature"). By performing a unified global comparison of this list, the smallest value is selected as the global minimum. The operational role identifier and feature dimension name corresponding to this global minimum together constitute the global extreme value feature node. This global extreme value feature node represents the single dimension that deviates most significantly from the optimal expectation among all operational capability feature dimensions within the current evaluation period, and its corresponding operational role.

[0183] For example, if the normalized global minimum value is 0.23, corresponding to the "closed-loop confirmation feature" dimension of the "pilot" role, then the global extreme value feature node is recorded as (pilot, closed-loop confirmation feature).

[0184] Through the multi-indicator comparison model described above, this module achieves unified quantitative evaluation across operational roles and dimensions, avoiding the local optimal perspective caused by analyzing only within a single role or dimension.

[0185] (ii) Feature Space Fusion

[0186] The feature node identification module concatenates the aforementioned global extreme value feature nodes (in the form of one-hot encoding or identifiers) with the cooperative quantization features output by the cooperative state quantization module 4 into a unified feature vector. The cooperative quantization features include at least:

[0187] Shared cognitive metrics (state alignment rate) for combinations of three operating roles;

[0188] The cooperative coupling coefficient for all two-way operation role pairs;

[0189] Optionally, the feature vector may also include original interaction sub-indicators for each operational role pair, including but not limited to: interaction frequency, confirmation rate, response latency, and handover integrity rate. These original sub-indicators are the basic data for calculating the collaborative coupling coefficient, and directly incorporating them into the feature space can further improve the model's ability to represent the collaborative state in a fine-grained manner. Through this fusion, the feature node identification module constructs a joint feature space that simultaneously describes the extreme nodes of individual operational capabilities and the quality of multi-role collaborative interactions.

[0190] (III) Risk prediction and contribution ranking of interpretable machine learning models

[0191] The feature node identification module uses a gradient boosting tree model as the interpretable machine learning model and integrates the Tree SHAP algorithm for interpreting feature contribution. The basic principles of the gradient boosting tree model and the Tree SHAP algorithm are well-known technologies in the field of machine learning, and this invention will not elaborate on their general mathematical derivations. The following only describes the specific configuration and usage of the model in this embodiment.

[0192] Model Architecture: Gradient boosting trees approximate the objective function by iteratively constructing an additive ensemble of multiple regression trees. Each new tree fits the negative gradient direction of the current model residual, using the anomaly level classification of the target operating system as the training label. Anomaly levels can be divided into multiple grades, such as 0-normal, 1-moderate anomaly, and 2-severe anomaly; the specific classification criteria can be set based on the severity of historical events. During model training, regularization terms (including the number of leaf nodes and the L2 norm of the leaf weight vectors) are used to control model complexity. Model Input and Output: The aforementioned fused feature vector (including global extreme value feature node encoding, shared cognitive metrics, cooperative coupling coefficients, etc.) is used as model input. The model output is the anomaly prediction value of the target operating system, ranging from 0 to 1, representing the probability of a collaboratively related anomaly event occurring under the current input feature combination. A preset threshold (e.g., 0.5) can be used to determine whether there is a significant anomaly risk.

[0193] Feature Contribution Ranking: After model training, the tree-structured Shapley algorithm is used to calculate the Shapley value (SHAP value) for each input feature. The Shapley value, based on the Shapley value theory in game theory, decomposes the model's predicted value into a baseline value and the sum of the contributions of each feature. The Shapley value of each feature quantifies its marginal contribution to the current prediction result: a positive value indicates that the feature increases the abnormal prediction value, while a negative value indicates that it decreases the abnormal prediction value. Features are sorted from largest to smallest according to the absolute value of their Shapley values, thus forming the contribution ranking.

[0194] Finally, the model outputs the anomaly prediction values ​​and the sorted Shapley values ​​corresponding to all input features. Both are used as the quantization input of the instruction generation module 6 to determine the priority processing order of the target feature dimensions.

[0195] In this embodiment, the target operating system refers to the entire system of multi-role collaborative operation being monitored (e.g., aviation operating system, industrial control system, etc.), which is the monitoring and prediction object of this invention. Anomalies refer to events or states that deviate from the preset normal operating state in the target operating system, including but not limited to: deviation of operator behavior from expected specifications, interruption or error in cross-role information transmission, incomplete task handover, and collaborative response timeout. The system anomaly prediction value is the probability assessment of the occurrence of the above-mentioned anomaly events in the target operating system within a certain time window in the future. In this embodiment, the feature node identification module 5 integrates the global extreme value feature nodes of individual operational capabilities with collaborative quantification features in the same feature space, enabling the machine learning model to identify coupling risk patterns that cannot be captured by a single profiling system. Specifically, when a certain operational capability feature dimension of an operator has an extreme value deviation (i.e., a global extreme value feature node), and this dimension happens to be located on the critical link of multi-role collaborative interaction, the machine learning model can learn the interaction between the individual feature and the collaborative feature, generating an additional predictive contribution beyond the sum of their independent contributions. The identification of this coupling risk depends on the joint input of individual and collaborative features, and a single-dimensional profiling system cannot perceive such risks at all due to the lack of a collaborative feature dimension.

[0196] In one specific embodiment, the input feature vector of the machine learning model includes the following components:

[0197] Quantitative scores for each operational role and each operational capability feature dimension: derived from the individual feature representation module, for example, the pilot's "closed-loop confirmation feature" score is 0.75, and the dispatcher's "timeliness of information transmission" score is 0.82, etc.

[0198] The role identifier and feature dimension name of the global extreme value feature node: for example (pilot, closed-loop confirmation feature), which represents the role and dimension corresponding to the global minimum value in the current evaluation period.

[0199] Shared cognitive metric of the three-party operation roles: namely, state alignment rate, with a value range of [0,1], reflecting the cognitive consistency among the task coordinator, the instruction issuer, and the operation executor.

[0200] The cooperative coupling coefficient of all two-party operation role pairs: The cooperative coupling coefficient of each operation role pair is used as an independent input feature.

[0201] Optionally, the input feature vector further includes a collaborative anomaly type identifier: this identifier is generated based on whether the collaborative coupling coefficient of each operational role pair is lower than a preset threshold. For example, for each operational role pair, if its collaborative coupling coefficient is lower than the anomaly threshold corresponding to that role pair (the anomaly threshold can be determined based on the statistical distribution of collaborative failure events of that role pair in historical data, such as taking the fifth percentile), then a discrete label is generated (e.g., 1 indicates the existence of a collaborative anomaly, and 0 indicates no anomaly). The above discrete labels can be concatenated into a binary vector as an additional feature group input model. Alternatively, for global collaborative anomalies, only a general identifier can be generated, which is set to 1 when the collaborative coupling coefficient of any operational role pair is lower than its threshold.

[0202] Through the above methods, the model can explicitly utilize collaborative risk indication information extracted based on prior knowledge or statistical regularities, further improving the accuracy of risk prediction. VI. Instruction Generation Module

[0203] The instruction generation module 6 is connected to the feature node recognition module 5 and the collaborative behavior knowledge base module 3 respectively, and is configured to: determine the target feature dimension based on the contribution ranking, retrieve observable behavior items that have a mapping relationship with the target feature dimension in the collaborative behavior knowledge base module, and automatically generate a structured behavior correction instruction set based on the retrieved observable behavior items and their hierarchical scoring criteria.

[0204] Specifically, this module performs the following operations in sequence:

[0205] (a) Determining the Dimensions of Target Features

[0206] The feature node identification module 5 outputs a contribution ranking (the absolute values ​​of the Shapley values ​​of each input feature are arranged from largest to smallest). At least one feature ranked highest is selected as the target feature dimension. This target feature dimension typically corresponds to the operational role and operational capability feature dimension indicated by the global extreme value feature node (e.g., "pilot-closed-loop confirmation feature"), but several features can also be selected based on actual needs.

[0207] (ii) Retrieval of observable behavioral items

[0208] In the collaborative behavior knowledge base module 3, observable behavior items that have a mapping relationship with the target feature dimension (i.e., the operation role type and the operation capability feature dimension name) are retrieved. The mapping relationship is pre-stored in the context adaptation layer (third layer) of the knowledge base, defining the correspondence between different operation roles and different operation capability feature dimensions and specific observable behavior items. For example, when the target feature dimension is "pilot - closed-loop confirmation feature", observable behavior items associated with this dimension are retrieved, such as "C3.4: actively clarifying ambiguous information".

[0209] (III) Generation of Structured Behavior Modification Instruction Set

[0210] Based on the retrieved observable behaviors and their rating criteria, a structured set of behavior correction instructions is automatically generated. This set of instructions includes the following components:

[0211] Behavior Correction Operation Sequence: For the behavior descriptions defined in the observable behavior items, generate specific operation steps or dialogue templates. For example, for the behavior item "proactively clarifying ambiguous information", the operation sequence is generated as follows: "1. Identify ambiguous elements after receiving information; 2. Use a standard follow-up question ('Please confirm...') within 5 seconds; 3. Record the confirmation reply."

[0212] Resource configuration parameters: Based on the type of the target feature dimension (the operational roles and capability dimensions involved) and the degree of deviation (reflected by the quantitative value of the corresponding dimension in the individual feature vector; the greater the deviation from the preset benchmark, the more severe the deviation), appropriate resource categories are automatically matched. Resource categories include, but are not limited to: full-mission simulators, fixed-program trainers, virtual reality collaborative training devices, computer-aided training terminals, and scenario-based training scripts. For example, if the target feature dimension involves multiple operational roles (such as collaborative behavior between pilots and dispatchers), virtual reality devices or simulators that support multi-role networked training will be prioritized; if only single-role operational behavior is involved, a single-seat trainer or computer-aided training terminal can be matched.

[0213] Task configuration parameters include the context definition and behavioral objectives required to perform the behavior correction task. The context definition is retrieved from the context adaptation layer (third layer) of the collaborative behavior knowledge base module, which calls the description of the operational scenario corresponding to the observable behavior item (e.g., "dispatchers use fuzzy expressions in collaborative decision-making for thunderstorm avoidance"). The behavioral objectives directly reference the observable feature description of the corresponding target level (e.g., "excellent demonstration") in the grading and scoring criteria for the observable behavior item. The task configuration parameters may also include suggested task execution durations (e.g., 30 minutes, 45 minutes, etc.).

[0214] (iv) Output and use of instruction sets

[0215] The generated structured behavior correction instruction set is output in a structured data format (such as JSON, XML, or tables), which can be directly parsed and executed by downstream training management systems or operators. The evaluation criteria in the instruction set directly reuse the hierarchical scoring criteria in the collaborative behavior knowledge base, ensuring the consistency between the behavior correction effect and the evaluation benchmark.

[0216] A specific example: If the target feature dimension is "Dispatcher-Intent Clarification Dimension" and the deviation is moderate, the system retrieves the associated observable behavior item "C2.2: Before making a detour suggestion, cite at least one specific meteorological or performance data as the basis for decision-making." The generated instruction set includes: behavior correction operation sequence - "Acquire meteorological data → Citify data → Make a suggestion"; resource configuration parameters - "Pilot-Dispatcher collaborative simulator environment"; task configuration parameters - "Subject name: Intent clarification training in detour collaborative decision-making; Scenario: Dispatcher gives ambiguous diversion suggestions during thunderstorm detours; Behavioral objective: Explicitly cite wind speed and fuel data in the suggestion; Suggestion duration: 45 minutes; Evaluation criteria: C2.2 behavior item five-level scoring anchor point."

[0217] Through the above configuration, this module realizes the automated mapping from the ranking of anomaly prediction contributions to structured, executable behavior correction instructions, and establishes a complete "assessment-intervention" technical link.

[0218] Furthermore, it also includes an interactive visual report module (not shown), connected to the instruction generation module, configured as follows:

[0219] The radar chart displays the individual feature vectors of each operational role, while also presenting quantified values ​​of multiple operational capability feature dimensions.

[0220] The trend of shared cognitive metric (state alignment rate) over time is shown using a line graph;

[0221] Use a barrel chart (or horizontal bar chart) to display the global extreme value feature nodes in the multi-indicator comparison model across operational roles and dimensions, highlighting the operational role identifier and feature dimension name corresponding to the node;

[0222] The behavior correction instruction set is displayed as a structured list, including the behavior correction operation sequence, resource configuration parameters, and task configuration parameters;

[0223] It provides interactive query and report generation functions, allowing users to filter and export visual results by operation role, time range, or task scenario.

[0224] With the above configuration, the interactive visualization report module presents the quantitative features, recognition results, and generated instruction sets output by each module in a graphical and structured form, making it easier for operators or training managers to intuitively understand the system output and perform subsequent operations.

[0225] Furthermore, it also includes a model self-optimization module (not shown), which is connected to the instruction generation module 6 and the collaborative behavior knowledge base module 3, respectively, and is configured to realize the system's self-calibration and parameter iteration. Specifically, this module performs the following operations:

[0226] (a) Feedback Data Collection

[0227] The system receives feedback data re-collected after the intervention operation is completed based on the behavior correction instruction set. This feedback data includes interaction data between operating roles, operational data, and training data re-collected in the same or equivalent simulated task environment. For ease of comparison, the feedback data is collected under the same conditions as the baseline data before intervention (e.g., task scenario, role configuration, data acquisition channel).

[0228] (ii) Quantification of behavioral changes

[0229] Based on the observable behavior items and their rating criteria defined in the Collaborative Behavior Knowledge Base Module 3, the system automatically compares the feedback data with the pre-intervention baseline data to calculate the score change for each observable behavior item. The score change is defined as the difference between the post-test score and the pre-test score (or the normalized value of the difference), with a positive value indicating improved behavior and a negative value indicating a decline.

[0230] (III) Judgment of intervention effect

[0231] Based on the score change magnitude, a quantified value of the correction effect of the behavior correction instruction set on a specific observable behavior item is calculated. The quantified value can be the weighted average or minimum of the score changes for each relevant observable behavior item. When this quantified value is lower than a preset effect threshold, the correction effect is deemed insufficient.

[0232] (iv) Automatic adjustment of mapping relationships and retrieval rules

[0233] When the quantified value of the correction effect is lower than the preset effect threshold, one or both of the following adjustments are triggered:

[0234] Adjust mapping relationships: Modify the association mapping relationship between observable behavior items and operational capability feature dimensions in the Collaborative Behavior Knowledge Base Module 3. For example, change the observable behavior item that the current target feature dimension was originally mapped to to another behavior item, or add / delete mapping entries.

[0235] Adjust retrieval matching rules: Modify the matching rules in instruction generation module 6 from contribution ranking to observable behavior item retrieval. For example, adjust the contribution ranking truncation number used to select target feature dimensions, or modify the similarity threshold used during retrieval.

[0236] Through the above configuration, this module realizes a complete technical closed loop of "quantitative evaluation → correction instructions → feedback collection → effect verification → parameter adjustment", enabling the system to self-calibrate based on actual feedback data during continuous operation and improve the accuracy of evaluation and correction.

[0237] Example 2

[0238] This embodiment uses a thunderstorm avoidance collaborative decision-making scenario as an example to demonstrate the complete operation process of the system from data acquisition to collaborative evaluation, feature node identification, command generation, and feedback verification. This scenario involves three operational roles: pilot, dispatcher, and controller, and is characterized by high time pressure and information density.

[0239] I. Data Collection Phase

[0240] The data acquisition module collects the following multi-source heterogeneous data through a dedicated data interface:

[0241] Air-to-ground communication recordings and transcripts (pilot-controller communications);

[0242] Data link communication messages (pilot-dispatcher messages, including detour suggestions, fuel data, and rerouting instructions, etc.);

[0243] Flight quality monitoring data (track deviation, altitude change rate, operational inputs, etc.);

[0244] Dispatcher's work log and release record.

[0245] The above data, after being timestamped, denoised, and standardized, is then used by downstream modules.

[0246] II. Individual Characteristic Representation Stage

[0247] The individual feature representation module extracts multi-dimensional operational capability feature quantification values ​​for the three operational roles based on a preset feature analysis framework (this embodiment adopts the ICAO-recognized operational capability framework). Taking a pilot as an example, their individual feature vector includes normalized scores for dimensions such as situational awareness, communication, teamwork, problem-solving and decision-making, and workload management.

[0248] III. Cooperative State Quantification Stage

[0249] The collaborative state quantification module retrieves observable behavior items matching the "thunderstorm avoidance collaborative decision-making" scenario from the collaborative behavior knowledge base. A graph neural network model is constructed using the three parties' roles as nodes and the collaborative coupling coefficient as edge weights to calculate the shared cognitive metric (state alignment rate). In this embodiment, the calculated state alignment rate of the three parties is 0.58 (range 0~1), which is lower than the preset threshold of 0.70, triggering a collaborative early warning signal.

[0250] IV. Feature Node Identification and Contribution Ranking

[0251] The feature node recognition module performs the following operations:

[0252] Construct a multi-indicator comparison model to identify global extreme value feature nodes: dispatcher Zhang San's "intention clarification" dimension (normalized score 0.56) and pilot Li Si's "closed-loop confirmation" dimension (normalized score 0.50).

[0253] The global extreme value feature nodes and collaborative quantization features (state alignment rate, collaborative coupling coefficient) are fused into a unified feature vector and input into the gradient boosting tree model (integrating the Tree SHAP algorithm).

[0254] Ranking of abnormal predicted values ​​and features in the model output: The pilot's "closed-loop confirmation" feature has the highest contribution (42%), followed by the dispatcher's "intention clarification" feature (35%), and the contribution of the interaction coupling term between the two is 23%.

[0255] V. Behavior Correction Instruction Generation Phase

[0256] The instruction generation module determines the target feature dimensions (pilot-closed-loop confirmation, dispatcher-intent clarification) based on contribution ranking, retrieves corresponding observable behavior items in the collaborative behavior knowledge base, and automatically generates a structured set of behavior correction instructions.

[0257] The generated instruction set for the target behavior item "C3.4 - Actively Clarify Ambiguous Information" of pilot Li Si includes:

[0258] Behavior correction operation sequence:

[0259] Identify trigger words: Initiate the clarification process when vague expressions such as "almost" or "should be" are heard;

[0260] Use standard follow-up questions: "What is the basis for your question? Can you confirm the specific figures?"

[0261] In a simulated environment, the training system or instructors repeatedly practice the role of dispatchers.

[0262] Resource configuration parameters:

[0263] Primary resource: Full mission simulator (supports pilot-dispatcher collaborative mode); Alternative resource: Virtual reality collaborative training equipment; Priority: Level 1 (contribution 42%).

[0264] Task configuration parameters:

[0265] Task Name: Training on Clarifying Fuzzy Information under High Pressure Environment; Scenario Definition: Collaborative Decision-Making for Flight Detours Amid Thunderstorms, Dispatchers Issue Fuzzy Statements (e.g., "It should be able to bypass it");

[0266] Behavioral objective: Initiate standard follow-up questioning techniques within 5 seconds of receiving ambiguous information;

[0267] Suggested duration: 30 minutes (including 3 rounds of repeated drills);

[0268] Evaluation criteria: The five-level rating scale of the C3.4 behavioral item is directly adopted, with a target score of ≥4.0.

[0269] For dispatcher Zhang San's target behavior item "C2.2 - Decision Basis Reference", the generated instruction set includes behavior correction operation sequence (referencing meteorological or performance data before proposing detour suggestions), resource configuration parameters (computer-aided training terminal + joint simulator environment), and task configuration parameters (subject duration 45 minutes, evaluation criteria referencing C2.2 level 5 scoring anchor points).

[0270] VI. Feedback Collection and Self-Optimization

[0271] After trainees complete training based on the behavior correction instruction set, the system collects feedback data again under the same simulated task environment. The model self-optimization module compares the baseline data before intervention with the feedback data and calculates the magnitude of score changes for each observable behavior item:

[0272] Pilot Li Si's C3.4 behavioral score improved from 2.3 to 3.8;

[0273] Dispatcher Zhang San's C2.2 behavioral score improved from 2.5 to 4.2.

[0274] The system marks C3.4 behavior items whose score changes are below the preset effect threshold (1.5 points improvement in this embodiment) as "continuous tracking items" and prioritizes retesting them in the next evaluation cycle; it marks qualified items as "qualified". The above process fully realizes the technical closed loop of "quantitative evaluation → instruction generation → feedback collection → effect verification".

[0275] Example 3

[0276] This embodiment uses a fault-holding release scenario as an example to demonstrate the quantification of collaborative status and the generation of behavioral correction instructions during the handover process among the three operational roles: maintenance engineer, dispatcher, and pilot. The core technical challenge in this scenario is that the completeness of fault handover information transmission across roles directly affects the accuracy of the pilot's understanding of operational restrictions.

[0277] The collaborative link in this scenario involves two serial handover stages:

[0278] Step 1: Aircraft Maintenance Engineer → Dispatcher (Information transmission for fault detection and retention decisions);

[0279] Step Two: Dispatcher → Pilot (Final communication of operational restrictions).

[0280] In the context adaptation layer of the collaborative behavior knowledge base, the following observable behavior items and their hierarchical scoring criteria are predefined for this scenario:

[0281] C1.1 (Structured Information Transmission): Information is transmitted using a four-part structure: "Fault Phenomenon → Evidence to be Retained → Operational Restrictions → Handling Suggestions";

[0282] C1.4 (Terminology Standardization): Strictly use Minimum Equipment List (MEL) entry coding;

[0283] C3.1 (Key Instruction Repetition): Repetition and confirmation of key data such as the entry number and operation restrictions;

[0284] C3.5 (Summary and Consensus): Three points will be summarized and confirmed at the end of the handover.

[0285] The data acquisition module collects the following multi-source heterogeneous data through a dedicated interface:

[0286] The electronic work card system retains release records;

[0287] Recordings and transcripts of conversations between maintenance engineers and dispatchers;

[0288] Records of ACARS messages and data link communications between dispatchers and pilots.

[0289] The collaborative state quantization module quantifies the closed-loop acknowledgment rate (i.e., number of acknowledgment events / total number of sent events, based on C3.1 behavioral item scoring) of the two handover links. The calculation results are as follows:

[0290] Closed-loop confirmation rate for step one (aircraft maintenance → dispatch): 62%;

[0291] The closed-loop confirmation rate for stage two (dispatch → pilot) is 95%.

[0292] The data above indicates that information transmission attenuation is mainly concentrated in the first handover stage.

[0293] The feature node identification and contribution ranking module constructs a multi-index comparison model, placing the scores of each action item of each operational role in the two links into the same numerical space for comparison. The global extreme feature node is identified as follows: the aircraft maintenance engineer has the lowest normalized score (0.48) in the "C1.1 (structured information transmission)" dimension. After fusing the global extreme feature node with collaborative quantification features (closed-loop confirmation rate, collaborative coupling coefficient), the input is given to an interpretable machine learning model, and the output is a ranking of feature contributions. The ranking results show that the aircraft maintenance engineer has the highest contribution to the "C1.1" feature (51%), followed by the "C3.1" feature (29%).

[0294] The behavior correction instruction generation module determines the target feature dimension (aircraft maintenance engineer - structured information transmission) based on contribution ranking, retrieves the corresponding observable behavior item C1.1 from the collaborative behavior knowledge base, and automatically generates a set of structured behavior correction instructions:

[0295] Behavior correction operation sequence:

[0296] During the fault hold and release process, a standardized script is output in sequence, consisting of four paragraphs: "Fault Phenomenon", "Basis for Hold (MEL Entries)", "Operational Restrictions", and "Handling Suggestions".

[0297] After each segment of the script, proactively ask the recipient if they understand it;

[0298] After all four sections are completed, the recipient is asked to reiterate the key restrictions.

[0299] Resource configuration parameters:

[0300] Primary resource: Computer-aided training terminal (CBT); Alternative resource: Virtual reality standalone training equipment; Priority: Level 1 (contribution 51%).

[0301] Task configuration parameters:

[0302] Task Name: Training on the Four-Segment Information Transmission for MEL Retention and Release;

[0303] Scenario Definition: Includes three common scenarios: APU fault retention, hydraulic system leakage retention, and navigation equipment downgrade release; Behavioral Objective: Complete four-segment structured communication, with the receiver's paraphrasing accuracy ≥90%;

[0304] Suggested duration: 60 minutes (20 minutes for each scenario, including repeated drills);

[0305] Evaluation criteria: The five-level rating scale of the C1.1 behavioral item is directly adopted, with a target score of ≥4.0.

[0306] Feedback Collection and Self-Optimization

[0307] After the trainee completes training based on the behavior correction instruction set, the system collects the interaction data of the maintenance engineer in the same or equivalent hold-and-release simulation task. The model self-optimization module compares the closed-loop confirmation rate before and after intervention: the closed-loop confirmation rate of stage one increases from 62% to 89%. The system marks this behavior item as "met the standard" and records sub-behaviors whose score changes below a preset threshold during training (such as omissions in the "operational restrictions" section) as items to be tracked for priority retesting in the next evaluation cycle.

[0308] Through the above process, this embodiment fully demonstrates the technical closed loop of the system in the fault-preservation release scenario, from data acquisition, collaborative state quantification, feature node identification, behavior correction instruction generation to feedback self-optimization.

[0309] Example 4

[0310] The system provided in this embodiment can be deployed in an airline's operations control center, adopting a hybrid cloud architecture, and the specific configuration is as follows:

[0311] Edge computing layer: Data acquisition modules and data preprocessing functions are deployed at edge nodes (such as airport sites and operation control centers) to collect voice streams, data link messages and operating parameters in real time, perform timestamp alignment and preliminary cleaning, and reduce data transmission latency.

[0312] Private Cloud Layer: Core computing modules such as the individual feature representation module, collaborative behavior knowledge base module, collaborative state quantification module, feature node recognition module, and instruction generation module run on a private cloud server cluster, utilizing graphics processing units (GPUs) to accelerate graph neural network inference and hidden Markov model decoding.

[0313] Enterprise Data Platform Layer: The contribution ranking output by the feature node recognition module and the behavior correction instruction set generated by the instruction generation module are connected to the airline's training management system, simulator scheduling system, and personnel qualification system through API interfaces to realize the automatic push and execution status tracking of behavior correction instructions.

[0314] Visualization layer: The interactive visualization report module is implemented using web technology and supports access on PCs and mobile devices.

[0315] All system data is stored in a distributed database, employing data anonymization and access control mechanisms to ensure that sensitive data such as individual feature vectors, interaction records, and abnormal event tags are accessible only to authorized users. The system interfaces with existing external systems such as flight quality monitoring systems, operation control systems, and training management systems through standard data interfaces (e.g., RESTful APIs, message queues) to achieve data sharing and business collaboration.

[0316] At the general deployment level, the aforementioned modules (data acquisition module 1, individual feature representation module 2, collaborative behavior knowledge base module 3, collaborative state quantification module 4, feature node identification module 5, instruction generation module 6, model self-optimization module 7, and an optional visualization report module) are deployed on a computing device containing at least one processor and memory. The modules are connected via data interfaces, with the output data of one module serving as the input data for the next, collectively forming a complete data processing chain from multi-source heterogeneous data acquisition, collaborative state quantification, feature node identification, behavior correction instruction generation to model self-optimization. The computing device can be a single server, a distributed computing cluster, or a cloud computing platform. Executable instructions are stored in the memory, and the processor executes these instructions to implement the functions configured in each module. The data interface can be a shared memory channel, message queue, RESTful API, or dedicated data bus; the specific implementation method is determined based on the deployment environment.

[0317] The hybrid cloud architecture presented in this embodiment is a preferred deployment method, but the scope of protection of this invention is not limited thereto. Any computing device and communication method that can realize the functions of each module and maintain the integrity of the data link falls within the scope of protection of this invention.

[0318] Example 5

[0319] This embodiment provides a cross-role collaborative quantization and instruction method for multi-source data fusion, which can be implemented based on the system described in Embodiment 1. Figure 2 As shown, the method includes the following steps:

[0320] S100: Multi-source heterogeneous data acquisition

[0321] Multi-source heterogeneous data from multiple operational roles is collected through a dedicated data interface. This data includes at least runtime data, training data, and interaction data between operational roles. For specific data types and collection methods, please refer to the description of data collection module 1 in Example 1.

[0322] S200: Individual Feature Vector Construction

[0323] Based on a pre-defined feature analysis framework, multi-dimensional operational capability features (such as operational accuracy, reaction time, and closed-loop confirmation features) of each operational role are extracted from the multi-source heterogeneous data to construct individual feature vectors. For specific extraction and normalization methods, please refer to the description of individual feature representation module 2 in Example 1.

[0324] S300: Collaborative Quantization Feature Calculation

[0325] A pre-built collaborative behavior knowledge base is invoked, which stores a standardized behavior assessment scale containing multiple observable behavior items and their rating criteria. Based on the individual feature vectors and observable behavior items from the knowledge base that match the current scene, collaborative quantification features between at least one pair of operational roles are calculated. These collaborative quantification features include: a shared cognitive metric (state alignment rate) for a three-way role combination and a collaborative coupling coefficient for a two-way role pair. For specific calculation methods (Hidden Markov Model, collaborative coupling coefficient formula, etc.), please refer to the description of collaborative state quantification module 4 in Example 1.

[0326] S400: Global Extreme Value Feature Node Recognition

[0327] Based on the individual feature vectors, a multi-index comparison model is constructed across operational roles and dimensions: the quantized values ​​of all feature dimensions of all operational roles are normalized and placed in the same numerical space for comparison. The operational role identifier and feature dimension name corresponding to the global minimum value are selected as the global extreme value feature node. For specific implementation, please refer to Part (I) of Feature Node Recognition Module 5 in Embodiment 1.

[0328] S500: Explainable Machine Learning Predictions and Contribution Ranking

[0329] The global extreme value feature nodes and the collaborative quantization features are fused into a unified feature vector, which is then input into an interpretable machine learning model (e.g., a gradient boosting tree model integrated with the Tree SHAP algorithm). The model outputs the anomaly prediction value of the target operating system and the Shapley value ranking (contribution ranking) of each input feature. For the specific training and inference process, please refer to Part (III) of Feature Node Recognition Module 5 in Example 1.

[0330] S600: Automatic generation of behavior correction instructions

[0331] Based on the contribution ranking, the target feature dimension is determined. Observable behavior items that have a mapping relationship with the target feature dimension are retrieved from the collaborative behavior knowledge base. Based on the hierarchical scoring criteria of the behavior item, a structured behavior correction instruction set is automatically generated, including behavior correction operation sequences, resource configuration parameters, and task configuration parameters. For specific generation logic, please refer to the description of the instruction generation module (6) in Example 1.

[0332] The above steps together constitute a complete methodology from data acquisition, feature extraction, collaborative quantization, node identification, predictive analysis to instruction generation. The specific technical details of each step in this embodiment have been fully disclosed in Embodiment 1 and will not be repeated here.

[0333] Compared with the prior art, the present invention has the following beneficial effects:

[0334] 1. Achieve unified collection and feature alignment of heterogeneous data across different roles.

[0335] This invention, by setting up a data acquisition module and utilizing a dedicated data interface, simultaneously collects operational data, training data, and interaction data between operational roles. It also performs timestamp alignment and semantic normalization on data from different sources, thereby breaking down the existing isolation between subsystems in terms of physical storage and logical models. This provides a complete and comparable data foundation for subsequent collaborative state quantification. Existing technologies can only access a single data source and cannot handle interactive data such as cross-position communication messages and handover records. This invention solves this data silo problem through the aforementioned technical means.

[0336] 2. For the first time, quantitative feature extraction of cross-role collaborative interaction process was achieved.

[0337] This invention establishes a collaborative state quantification module, which calculates collaborative quantification features (such as shared cognitive metrics and collaborative coupling coefficients) between at least one pair of operational roles based on individual feature vectors and observable behavioral items retrieved from a collaborative behavior knowledge base. Existing systems can only extract operational parameters or physiological signals of a single role and cannot quantify the quality of interaction between roles. The collaborative quantification features of this invention directly reflect collaborative interface characteristics such as information transmission integrity, response timeliness, and confirmation rate.

[0338] 3. Establish an automated mapping from exception detection to executable instructions.

[0339] This invention establishes a feature node identification module and an instruction generation module. The feature node identification module outputs a ranking of the contributions of each feature to anomaly prediction using an interpretable machine learning model. The instruction generation module retrieves corresponding observable behavior items from a collaborative behavior knowledge base based on this ranking and automatically generates a structured set of behavior correction instructions (including behavior correction operation sequences, resource configuration parameters, and task configuration parameters) based on its hierarchical scoring criteria. Existing technologies output anomaly detection results as high-level semantic labels (such as "communication timing disorder"), which rely on human experience to convert into low-level control instructions, leading to delays and subjective biases. This invention achieves automated mapping from quantitative analysis to executable instructions through the above modules, eliminating the manual conversion step.

[0340] 4. Improve the attribution accuracy and interpretability of anomaly prediction.

[0341] This invention fuses global extreme value feature nodes (the role and feature dimension corresponding to the minimum value obtained through cross-role and cross-dimensional comparison) with collaboratively quantized features into a unified feature vector. This vector is input into a gradient boosting tree model and integrated with the Tree SHAP algorithm, outputting a ranking of the Shapley values ​​for each feature. This ranking directly quantifies the marginal contribution of each feature to anomaly prediction. Compared to existing black-box models that only provide anomaly probabilities, this invention can accurately locate key features leading to anomalies (e.g., the "pilot-closed-loop confirmation feature" contributes 42%), enhancing the interpretability of prediction results and providing a quantitative basis for intervention.

[0342] 5. Possesses scenario expansion and self-optimization capabilities.

[0343] This invention's collaborative behavior knowledge base adopts a three-layer decoupled architecture (collaborative dimension layer, observable behavior item layer, and context adaptation layer). For new scenarios, only mapping rules need to be added to the context adaptation layer; no modification to the underlying behavior item library is required. Simultaneously, the model self-optimization module automatically compares the changes in observable behavior item scores based on post-intervention feedback data. When the change falls below a preset threshold, it triggers automatic adjustments to mapping relationships or retrieval rules. Existing systems have fixed evaluation frameworks and cannot iterate with new data. This invention, through the aforementioned architecture and modules, achieves system self-adaptation and self-optimization to operating scenarios, improving long-term accuracy and adaptability.

[0344] This invention also provides an electronic device, including: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being configured to perform the method described in this invention.

[0345] This invention also provides a computer-readable storage medium storing computer-executable instructions for performing the methods described in this invention.

[0346] It should be understood that the various forms of processes shown above can be used to reorder, add, or delete steps. For example, the steps described in this invention can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution disclosed in this invention can be achieved, and this is not limited herein.

[0347] The specific embodiments described above do not constitute a limitation on the scope of protection of this invention. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this invention should be included within the scope of protection of this invention.

Claims

1. A cross-role collaborative quantization and instruction system for multi-source data fusion, characterized in that, The system includes: The data acquisition module is configured to collect multi-source heterogeneous data from multiple operational roles through a dedicated data interface. The multi-source heterogeneous data includes operational data, training data, and interaction data between operational roles. The operational data includes at least one of the following: status monitoring parameters related to the execution process of the operational task, task timing records, and resource configuration records. The training data includes at least one of the following: operational behavior records in a simulated environment and capability assessment results. The individual feature representation module is configured to extract multi-dimensional operational capability features of each operational role from the multi-source heterogeneous data based on a preset feature analysis framework, and construct an individual feature vector. The collaborative behavior knowledge base module is configured to store a standardized behavior assessment scale for collaborative scenarios involving multiple operational roles. The standardized behavior assessment scale includes multiple observable behavior items and their corresponding graded scoring criteria. The collaborative state quantization module is configured to receive the individual feature vector output by the individual feature representation module, and call the observable behavior items that match the current scene from the collaborative behavior knowledge base module, and calculate the collaborative quantization features between at least one pair of operation roles based on the two. The feature node identification module is configured to construct a multi-index comparison model across operational roles and dimensions based on the individual feature vectors, identify global extreme value feature nodes, and input the global extreme value feature nodes and the collaborative quantization features into an interpretable machine learning model, and output the contribution ranking of each feature to the anomaly prediction of the target operating system. The instruction generation module is configured to determine the target feature dimension based on the contribution ranking, retrieve observable behavior items that have a mapping relationship with the target feature dimension from the collaborative behavior knowledge base module, and automatically generate a structured behavior correction instruction set based on the retrieved observable behavior items and their hierarchical scoring criteria.

2. The system according to claim 1, characterized in that, The data exchanged between the operational roles includes at least one of the following: voice communication records and corresponding text transcriptions, structured data link messages, internal call audio records, and electronic records of task handover.

3. The system according to claim 1, characterized in that, The standardized behavior assessment scale adopts a three-layer structure: the first layer defines multiple collaborative behavior dimensions that are common to multiple operational roles; The second layer stores the multiple observable behavior items and their hierarchical scoring criteria under each collaborative behavior dimension; The third layer of storage context adaptation mapping rules defines the correspondence between the observable behavior items and specific operation role pairs under different operating scenarios.

4. The system according to claim 1, characterized in that, The specific configuration of the collaborative state quantification module is as follows: For operation role combinations that include three parties—task coordinator, instruction issuer, and operation executor—a hidden Markov model is used to jointly model the three-party interaction behavior sequence, and the state alignment rate is calculated as a shared cognitive metric; for operation role pairs that only include two parties, the collaborative coupling coefficient is calculated based on interaction frequency, confirmation rate, response latency, and handover integrity rate.

5. The system according to claim 4, characterized in that, The collaborative state quantization module is further configured to: construct a graph neural network model with each operation role as a node and the collaborative coupling coefficient as the edge weight, update node features through graph convolution operations, and output the global collaborative efficiency coefficient after global pooling. The global cooperative efficiency coefficient is used to adjust the parameters of the hidden state transition probability matrix in the hidden Markov model.

6. The system according to claim 1, characterized in that, The interpretable machine learning model in the feature node identification module is a gradient boosting tree model, and the tree-based Shapley algorithm is used to calculate the Shapley value of each input feature as the contribution ranking.

7. The system according to claim 1, characterized in that, The structured behavior correction instruction set includes: a behavior correction operation sequence corresponding to the observable behavior item, resource configuration parameters, and task configuration parameters; wherein, the resource configuration parameters automatically match resource categories according to the type and deviation degree of the target feature dimension; the task configuration parameters include the context definition and behavior objectives required for task execution, and the measurement criteria corresponding to the behavior objectives directly reference the hierarchical scoring criteria of the corresponding observable behavior item in the collaborative behavior knowledge base module.

8. The system according to claim 1, characterized in that, The feature node identification module constructs a multi-index comparison model across operation roles and dimensions, which is used to place the quantified values ​​corresponding to all operation capability feature dimensions of all operation roles in the same numerical space for unified comparison, and selects the operation role identifier and feature dimension name corresponding to the global minimum value from the comparison results as the global extreme value feature node.

9. The system according to claim 1, characterized in that, It also includes a model self-optimization module, which is configured to: receive feedback data re-collected after the intervention operation is completed based on the behavior correction instruction set; compare the feedback data with the baseline data before intervention based on the observable behavior items defined in the collaborative behavior knowledge base module; and calculate the score change magnitude of each observable behavior item. When the score change magnitude is lower than a preset threshold, it automatically adjusts the mapping relationship between the observable behavior items and the operational ability feature dimensions in the collaborative behavior knowledge base module, or adjusts the retrieval matching rules in the instruction generation module.

10. A cross-role collaborative quantization and instruction method for multi-source data fusion, characterized in that, Includes the following steps: Multi-source heterogeneous data from multiple operational roles is collected through a dedicated data interface. The multi-source heterogeneous data includes operational data, training data, and interaction data between operational roles. The operational data includes at least one of the following: status monitoring parameters related to the execution process of operational tasks, task timing records, and resource configuration records. The training data includes at least one of the following: operational behavior records in a simulated environment and capability assessment results. Based on a preset feature analysis framework, multi-dimensional operational capability features of each operational role are extracted from the multi-source heterogeneous data to construct individual feature vectors. A pre-built collaborative behavior knowledge base is invoked, which stores a standardized behavior assessment scale containing multiple observable behavior items and their graded scoring criteria; based on the individual feature vector and the observable behavior items that match the current scene and are invoked from the collaborative behavior knowledge base, at least one set of collaborative quantitative features between operation role pairs is calculated. Based on the individual feature vectors, a multi-indicator comparison model is constructed across operational roles and dimensions to identify global extreme value feature nodes; The global extreme value feature node and the collaborative quantization feature are used as input features and fed into an interpretable machine learning model. The model outputs a ranking of the contributions of each feature to the anomaly prediction of the target operating system. Based on the contribution ranking, the target feature dimension is determined. Observable behavior items that have a mapping relationship with the target feature dimension are retrieved from the collaborative behavior knowledge base. Based on the retrieved observable behavior items and their hierarchical scoring criteria, a structured behavior correction instruction set is automatically generated.

Images