Intelligent airport-oriented flight operation delay timing causal reasoning and intelligent processing method

CN122572690BActive Publication Date: 2026-10-13ZHEJIANG AIRPORT DIGITAL TECH CO LTD +3
View PDF 5 Cites 0 Cited by

Patent Information

Application Number
CN202611059473.8
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2026-07-16
Publication Date
2026-10-13
Estimated Expiration
2046-07-16

AI Technical Summary

Technical Problem

[0013]为了解决背景技术中存在的问题,本发明提供了一种面向智慧机场的航班运行延误时序因果推理与智能处理方法,解决了现有技术中保障节点级延误预测能力不足、因果分析与预测任务割裂、缺乏智能闭环决策以及模型适应性差等技术问题

Benefits of technology

[0052]1、实现节点级延误预测,解决传统预测粒度不足的问题。本发明通过时序因果模型将延误分析下沉至具体作业节点,基于多尺度时空特征输出各节点的延误时长预测值,为延误前置干预提供精细化的量化依据。

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN122572690B_ABST
    Figure CN122572690B_ABST
Patent Text Reader

Abstract

The application discloses a kind of intelligent airport-oriented flight operation delay timing causal inference and intelligent processing method.Method includes to historical operation data preprocessing, obtain data subset and cluster center;Train timing causal model and extract causal matrix;Configure delay root cause node information module and build scheduling strategy dataset;According to the scheduling strategy dataset, the scheduling strategy model is trained by reinforcement learning;Finally, the intelligent scheduling decision of real-time operation data is realized by using delay root cause node information module and the scheduling strategy model trained finally.The application forms the real-time closed loop of "node-level causal prediction-Root cause accurate score-intelligent dynamic decision" by delay root cause node information module and scheduling strategy model, and generates multiple and adaptive scheduling decision results under the premise of strictly observing physical constraints by using improved DPO loss function for reinforcement training.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention belongs to the field of smart airport operation management and artificial intelligence technology, specifically involving a method for temporal causal reasoning and intelligent processing of flight operation delays for smart airports. Background Technology

[0002] With the rapid development of the civil aviation industry, the complexity of airport operation systems is constantly increasing. Flights typically undergo multiple closely linked operational stages during the ground handling phase, requiring coordinated scheduling of tasks such as docking at aircraft gates, passenger disembarkation, baggage and cargo unloading, aircraft inspection, and cleaning operations. Some of these stages have strict temporal dependencies; a delay at one stage can gradually propagate through the handling process to subsequent stages, ultimately leading to flight delays or even a decrease in airport operational efficiency.

[0003] Currently, several patented technologies have been proposed for flight delay analysis and prediction. For example, CN110659773A discloses a deep learning-based flight delay prediction method that trains a neural network model using historical flight data to predict the duration of flight departure delays. This type of method primarily focuses on predicting overall delays and fails to refine the estimation of delay probabilities for each support node, making it difficult to support sophisticated ground dispatching decisions.

[0004] CN121212904A discloses a method and system for assessing and analyzing airport operation support capabilities. By intelligently predicting the number of arriving aircraft and passenger traffic, it dynamically adjusts airport operation and maintenance resources, achieving an overall assessment and optimized resource allocation of airport support capabilities from a macro perspective. However, this method focuses on macro-level resource prediction and allocation, failing to delve into the delay analysis and prediction of specific operational nodes in the flight ground support process. It lacks modeling of the temporal dependencies between support nodes, making it difficult to identify the propagation path and cascading effects of delays in the support process. Furthermore, it does not establish a closed-loop mechanism from delay analysis to real-time operation and maintenance decision-making. Its resource adjustment strategy is primarily based on macro-level prediction rather than real-time response to the current operational situation.

[0005] CN120633951A discloses a flight delay prediction method based on multimodal data fusion and large model fine-tuning. This method constructs a state-space model to characterize the dynamic characteristics of delays, fine-tunes the Qwen large model using the LoRA method, combines real-time data for delay prediction, and finally weights and fuses the large model's prediction results with the state-space model's estimates, effectively improving the accuracy and reliability of the prediction. However, the LoRA fine-tuning in this method only adjusts a small number of parameters of the large model. While this reduces computational costs, it also limits the large model's ability to fully learn complex delay patterns and multi-stage correlation features, making it difficult to accurately capture the cascading propagation effect of delays between support process nodes. Furthermore, this method focuses on predicting delay outcomes and lacks a real-time closed-loop control mechanism from prediction results to specific support operation nodes, failing to directly translate delay warnings into targeted operation and maintenance scheduling instructions.

[0006] The existing technology has the following main pain points and shortcomings:

[0007] 1. Insufficient prediction granularity: Existing methods mostly focus on predicting the final takeoff delay, failing to delve into the delay probability estimation at the support node level, resulting in the inability to carry out targeted interventions before delays occur.

[0008] 2. Separation of causality and prediction: Existing causal analysis methods are mainly used for post-event root cause localization, making it difficult to apply the causal structure in real time to the prediction task of node delays, resulting in a separation between "analysis" and "prediction".

[0009] 3. Superficial decision-making basis makes it difficult to identify root causes: While existing delay prediction methods based on large models can output delay results, their decision-making process relies solely on the statistical correlation of delay data, lacking an understanding of the inherent causal mechanisms of the assurance process. For example, although a certain preceding node is delayed, the final takeoff time can still be on time due to buffering or resource allocation in subsequent nodes. If the large model has not been trained with causal reinforcement, it is difficult to identify the real key root causes through the appearance of delays, easily misjudging superficial delays as actual risks, leading to decision-making biases and resource misallocation.

[0010] 4. Lack of intelligent closed-loop decision-making capability: Existing technologies only provide delay analysis or prediction results, failing to form a closed-loop mechanism from root cause identification to operation and maintenance decision-making, and unable to combine the reasoning capabilities of large models to generate executable operation strategies.

[0011] 5. Static models have poor adaptability: Most models are trained offline, making it difficult to dynamically adapt to changes in airport operating conditions and to develop timely response strategies for real-time airport operation and maintenance data.

[0012] Therefore, it is necessary to propose a method that can integrate the timing dependencies of guarantee nodes, real-time operation and maintenance data and planned time to achieve node-level delay prediction, and generate intelligent operation and maintenance decisions through large-scale model reinforcement learning, so as to assist airport managers in making forward-looking operation optimization and scheduling adjustments. Summary of the Invention

[0013] To address the problems existing in the background technology, this invention provides a time-series causal reasoning and intelligent processing method for flight operation delays in smart airports, which solves the technical problems in the prior art such as insufficient ability to predict delays at the node level, separation of causal analysis and prediction tasks, lack of intelligent closed-loop decision-making, and poor model adaptability.

[0014] The technical solution adopted in this invention is:

[0015] I. A method for temporal causal reasoning and intelligent processing of flight operation delays for smart airports:

[0016] S1. Obtain historical airport operation and maintenance data and perform cleaning, feature processing and clustering processing in sequence to obtain several data subsets and each cluster center vector.

[0017] S2. Construct a time-series causal model. Train the corresponding time-series causal model for each data subset to obtain several trained time-series causal models, and extract the causal matrix corresponding to each time-series causal model.

[0018] S3. Configure the delay root cause node information module using the cluster center vectors, the trained time-series causal model and the corresponding causal matrix, and construct the scheduling strategy dataset based on historical operation and maintenance data and the delay root cause node information module.

[0019] S4. Initialize the scheduling strategy model. Train the model using DPO-based reinforcement learning with the constructed scheduling strategy dataset to obtain the completed reinforcement training scheduling strategy model and complete local deployment.

[0020] S5. Intelligent scheduling decisions are made based on real-time airport operation and maintenance data through the delay root cause node information module and the scheduling strategy model completed by reinforcement training.

[0021] Step S1 specifically involves:

[0022] S11. Obtain historical operation and maintenance data of the airport and perform cleaning and feature processing in sequence to obtain processed historical operation and maintenance data.

[0023] The historical operation and maintenance data includes ground support data and scheduling decision data for each flight in the airport for several consecutive historical periods; the ground support data includes data from multiple operational nodes.

[0024] S12. Extract the delay time corresponding to each operation node from the processed flight ground support data. Use the flight ground support data as the sample input and the delay time of each operation node as the sample output label to construct a single data sample and obtain the delay root cause dataset.

[0025] S13. Use K-Means to cluster the delay root cause dataset to obtain several data subsets and the cluster center vector corresponding to each data subset.

[0026] The temporal causal model in step S2 includes an embedding coding layer, a backbone network, and a prediction head. Flight ground support data in the data subset is input into the embedding coding layer for processing to obtain multi-scale features. The multi-scale features are input into the backbone network, which includes multiple cascaded Decoder-Only modules. The last Decoder-Only module outputs multi-dimensional features. The multi-dimensional features are then processed by the prediction head to output the delay time of each operation node in the flight ground support data.

[0027] The specific processing flow of the embedded coding layer is as follows: the flight ground support data in the data subset is first processed by three parallel one-dimensional convolutions to obtain different feature maps, and the three feature maps are concatenated to obtain transform fusion features; the input data is processed by RoPE rotation position coding to obtain position coding features, the time intervals in the input data are encoded to obtain time coding features, and the position coding features and time coding features are gated and fused to obtain spatiotemporal coding features; the spatiotemporal coding features and transform fusion features are added to obtain multi-scale features.

[0028] The specific processing flow of each Decoder-Only module is as follows: the module input is first normalized by layers and then fed into the multi-head mask self-attention module to obtain self-attention features. The self-attention features and the module input residuals are concatenated to obtain preliminary fusion features. The preliminary fusion features are then passed through layers normalization and feedforward layers to obtain feedforward features. The feedforward features and the module input are concatenated by residuals and used as the output of the Decoder-Only module.

[0029] For each trained temporal causal model, the weights of all attention points in the model are taken to construct the corresponding causal matrix.

[0030] The scheduling strategy dataset in step S3 includes a historical scheduling strategy dataset and a strong model distillation dataset. The specific process of constructing the historical scheduling strategy dataset is as follows: acquire scheduling decision data and corresponding historical operation and maintenance data, use the historical operation and maintenance data as the data to be inferred, input the data to be inferred into the delay root cause node information module to output the delay root cause node information of each job node; use the historical operation and maintenance data and the delay root cause node information of each job node as sample input, use the historical scheduling decision data as the sample output label, and construct a historical scheduling strategy data sample. The specific process of constructing the strong model distillation dataset is as follows: Obtain historical operation and maintenance data for which no corresponding scheduling decision data exists. Use this historical operation and maintenance data as the data to be inferred. Input this data into the delay root cause node information module to output the delay root cause node information for each job node. Input the historical operation and maintenance data and the delay root cause node information for each job node into the strong model to output the scheduling decision results. Apply a phased preference labeling strategy to the output scheduling decision results, retaining the preferred labels as accepted scheduling decision results. Use the retained scheduling decision results as sample output labels, and use the corresponding historical operation and maintenance data and the delay root cause node information for each job node as sample inputs to construct the strong model distillation data samples.

[0031] The phased preference labeling strategy includes two phases. In the first phase, airport scheduling experts manually label the scheduling decision results output by the strong model as either accepting or rejecting them. In the second phase, a preference model is used to label the scheduling decision results output by the strong model.

[0032] The preference model is trained by the following steps: taking the scheduling decision results of the first stage as input and the manual preference labeling as output to construct seed samples, training the preference model based on the seed samples, and thus obtaining the preference model for preference labeling in the second stage.

[0033] The specific processing flow of the delay root cause node information module is as follows:

[0034] D1. Match the data to be inferred with each cluster center vector to obtain the most matching cluster center vector. Input the data to be inferred into the time-series causal model corresponding to the most matching cluster center vector for processing to obtain the causal matrix and the delay time of each task node.

[0035] D2. The delay amplification factor is calculated by combining the delay time of each operation node with the topological path in the process and the scale of downstream reachable nodes of each node. Based on the preset propagation scale parameters and Weibull shape parameters of each node, and combined with the risk recursion relationship of downstream immediate process nodes, the propagation risk intensity of each operation node is calculated by the fixed point iteration method.

[0036] D3. Calculate the delay risk score for each work node based on the causal matrix, the node propagation risk intensity of each work node, and the corresponding delay amplification factor.

[0037] D4. Summarize the propagation risk intensity, delay amplification factor, and delay risk score of each node to generate the delay root cause node information required for the scheduling model input.

[0038] The delay risk score for each task node in step D3 is set according to the following formula:

[0039] ;

[0040] ;

[0041] ;

[0042] in, , All are node indexes; For indexing time-series causal models; For nodes Delay risk score; This represents the total number of task nodes. For the first The nodes in the causal matrix corresponding to each time-series causal model For nodes The causal strength value; It is a nonlinear exponent of causal strength; For nodes With nodes The number of edges in the shortest directed path in the guarantee process; The topological distance Gaussian decay width parameter; For nodes The intensity of the risk of transmission; For propagation-amplification coupling coefficient; For nodes The delay amplification factor; and They are nodes and nodes The delay time; For nodes The propagation scale parameter; It is the gamma function; For nodes The Weibull shape parameters; The cascade coupling coefficient; For nodes The intensity of the risk of transmission; For nodes The set of downstream adjacent nodes; The causal strength coupling index; It is also a node index; and They are slave nodes and nodes The set of all directed paths to the final node; For nodes The total number of downstream nodes; It is a directed path; For directed paths The edges in; For directed paths The number of sides; The path length decay exponent; For the first In the causal matrix corresponding to each time-series causal model, the edges The causal strength value.

[0043] The loss function used for reinforcement training of the scheduling strategy model is set according to the following formula:

[0044] ;

[0045] ;

[0046] ;

[0047] ;

[0048] in, These are the parameters of the network to be trained; This is the total loss function; For DPO loss; To encourage exploration of dynamic weighting coefficients for the entropy loss term; This is the current training round; Total number of training rounds; This represents the difference in total loss between two adjacent training rounds; This is the initial value of the entropy adjustment coefficient; This represents the weight decay coefficient for the step dimension; This is the adjustment factor for the dimension of loss change; This is the sensitivity coefficient to changes in loss; To encourage the exploration of entropy loss; This is the entropy adjustment coefficient that is dynamically updated as the training process progresses. This represents the set of feasible scheduling strategies under the current scheduling state. Information on the root causes of delays is input into the model; To output the scheduling decision results; Input delay root cause node information In this case, the scheduling strategy model outputs the scheduling decision result. The probability value; It is a natural constant.

[0049] II. A computer device, comprising a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the above-described method.

[0050] 3. A computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of the above method.

[0051] The beneficial effects of this invention are:

[0052] 1. Achieve node-level delay prediction, solving the problem of insufficient granularity in traditional predictions. This invention uses a time-series causal model to analyze delays down to specific operational nodes, and outputs predicted delay durations for each node based on multi-scale spatiotemporal characteristics, providing a refined quantitative basis for proactive delay intervention.

[0053] 2. This invention achieves a fusion of prediction and causal analysis outputs, resolving the issue of their separation. Based on a time-series prediction model, this invention extracts the attention weights of the Decoder-Only module, and constructs a causal matrix after masking constraints and spectral truncation denoising. This allows the model to simultaneously reveal the strength of causal relationships in delay propagation between nodes while outputting delay prediction results, thus achieving an integrated output of prediction results and causal explanations.

[0054] 3. Achieve accurate root cause identification and solve the problem of superficial decision-making basis. This invention integrates causal strength, node propagation risk, and delay amplification effect to construct a multi-dimensional delay risk scoring system, which can penetrate the surface to identify key root cause nodes with cascading propagation effects, avoiding decision-making bias and resource misallocation caused by misjudging superficial delays as core risks.

[0055] 4. This invention establishes intelligent closed-loop decision-making capabilities, addressing the lack of a closed-loop mechanism from root cause identification to operational decision-making. It constructs a real-time closed loop of "node-level causal prediction - accurate root cause scoring - intelligent dynamic decision-making" through a delay root cause node information module and a scheduling strategy model. An improved DPO loss function is used for reinforcement training, generating diverse scheduling decision results that align with business preferences while strictly adhering to physical operational constraints.

[0056] 5. Enhance the model's scenario adaptability and generalization ability, addressing the problem of insufficient adaptability of a single static model. This invention trains multiple time-series causal models by clustering and segmenting scenarios to adapt to different types of delay propagation patterns; simultaneously, real-time operation and maintenance data and delay root cause node information are jointly input into the scheduling model to dynamically respond to changes in operating conditions such as aircraft location and weather. Furthermore, through strong model distillation and a phased preference labeling strategy, a high-quality training dataset can be constructed with limited labeling costs, further improving the model's generalization performance and deployment adaptability. Attached Figure Description

[0057] Figure 1 This is a flowchart of the method of the present invention.

[0058] Figure 2 This is a complete work node diagram of the guarantee process in this embodiment. Detailed Implementation

[0059] The present invention will now be described in more detail with reference to the accompanying drawings and embodiments. However, the present invention is not limited thereto. For those skilled in the art, several improvements and modifications can be made without departing from the principles of the present invention, and these improvements and modifications are also considered to be within the scope of protection of the present invention. Contents not described in detail in this specification are prior art known to those skilled in the art.

[0060] In this embodiment, historical operation and maintenance data from a major hub airport over 12 consecutive months was selected as the basic data source. Considering the significant differences in flight ground support procedures and operation durations among different aircraft types, to ensure the consistency and stability of model building, this embodiment selected historical operation and maintenance data corresponding to the two most prevalent mainstream aircraft types (narrow-gauge aircraft / wide-gauge aircraft) for implementation.

[0061] like Figure 1 As shown, the flight delay time-series causal reasoning and intelligent processing method of this embodiment is implemented according to the following steps:

[0062] S1. Obtain historical airport operation and maintenance data and perform cleaning, feature processing and clustering processing in sequence to obtain several data subsets and each cluster center vector.

[0063] S11. Obtain historical operation and maintenance data of the airport and perform cleaning and feature processing in sequence to obtain processed historical operation and maintenance data.

[0064] like Figure 2 As shown, the historical operation and maintenance data includes ground support data and scheduling decision data for each flight in the airport for several consecutive historical periods; the ground support data includes data for a total of 9 operation nodes, including docking at the bridge, passenger disembarkation, baggage and cargo unloading, aircraft inspection, cleaning operations, baggage and cargo loading, passenger boarding, pushback and takeoff.

[0065] Specifically, each job node's data includes the following 16 status indicators, and the description and quantified encoding of each status feature are shown in Table 1 below:

[0066] Table 1:

[0067]

[0068] Specifically, in the historical operation and maintenance data, some flight ground support data has corresponding scheduling decision data, while other flight ground support data does not have corresponding scheduling decision data.

[0069] The cleaning process specifically includes: using time series interpolation or statistical mean of similar flights to complete missing timestamps in historical operation and maintenance data, and filtering out abnormal data. Feature processing specifically includes: standardizing continuous value data in historical operation and maintenance data to meet the requirements of zero mean and unit variance distribution; and converting discrete value data into numerical representation through one-hot encoding.

[0070] Furthermore, the processed data from a single sample is organized into a three-dimensional tensor [T, N, F] as standardized temporal input structure data, where T is the length of the flight ground support data sequence contained in a single sample, N is the number of operational nodes, and F is the dimension of the state indicator in the data of each operational node; the input is organized in batches during model training. In this embodiment, T is set to 16, indicating that a single sample contains flight ground support data from 16 consecutive time points; that is, the training batch size is set to 16, consistent with the model training parameters.

[0071] Each job node's data includes the corresponding delay time.

[0072] S12. Extract the delay time of each operation node from the processed flight ground support data. Use the flight ground support data as the sample input and the delay time of each operation node as the sample output label to construct a single data sample and obtain the delay root cause dataset.

[0073] S13. Use K-Means to cluster the delay root cause dataset to obtain several data subsets and the cluster center vector corresponding to each data subset.

[0074] Specifically, using data samples in the delay root cause dataset as clustering units, K-Means clustering is performed on the delay root cause dataset to obtain several data subsets, each of which includes several flight ground support data.

[0075] In this embodiment, K-Means clustering is used to generate 5 data subsets.

[0076] S2. Construct a time-series causal model. Train the corresponding time-series causal model for each data subset to obtain several trained time-series causal models, and extract the causal matrix corresponding to each time-series causal model.

[0077] The temporal causal model in step S2 includes an embedding coding layer, a backbone network, and a prediction head. Flight ground support data in the data subset is input into the embedding coding layer for processing to obtain multi-scale features. The multi-scale features are then input into the backbone network, which includes multiple cascaded Decoder-Only modules. The last Decoder-Only module outputs multi-dimensional features, which are then processed by the prediction head to output the delay time of each operation node in the flight ground support data.

[0078] The specific processing flow of the embedded coding layer is as follows: the flight ground support data in the data subset is first processed by three parallel one-dimensional convolutions (1D CNN) with different scale convolution kernels to obtain different feature maps. The three feature maps are then concatenated to obtain transform fusion features. The input data is processed using RoPE rotation position coding to obtain position coding features. The time intervals in the input data are encoded to obtain time coding features. The position coding features and time coding features are then gated and fused to obtain spatiotemporal coding features. The spatiotemporal coding features are added to the transform fusion features to obtain multi-scale features. Specifically, RoPE rotation position coding uses the RoPE coding module in the Qwen3 model architecture.

[0079] In this embodiment, the kernels of the three one-dimensional convolutions are set to 3, 5, and 7, respectively, to extract the changing trend features of each operation node in the flight ground support data in the short and medium-to-long term, and then generate transform fusion features.

[0080] The processing flow of each Decoder-Only module is as follows: The input to the Decoder-Only module is first normalized through a first-layer normalization process, and then fed into a multi-head mask self-attention module to obtain self-attention features. These self-attention features are then residually concatenated with the input to the Decoder-Only module to obtain preliminary fused features. These preliminary fused features are then sequentially normalized through a second-layer normalization process and a feedforward layer to obtain feedforward features. These feedforward features are residually concatenated with the module input and used as the output of the Decoder-Only module. The input to the Decoder-Only module is the output of the previous layer / module.

[0081] Specifically, the feedforward layer uses a two-layer fully connected network, and the prediction head also uses a two-layer fully connected network. The AdamW optimizer is employed, with a learning rate of 1e-4, batch size of 16, epochs of 50, and MSE as the loss function. The training / validation / test sets are divided in a 7:1:2 ratio.

[0082] In this embodiment, the multi-head mask self-attention module adopts the 8-head self-attention of the decoder in the Transformer architecture.

[0083] For each trained temporal causal model, the corresponding causal matrix is ​​constructed based on the weights of all attention points in the model.

[0084] Specifically, for each trained temporal causal model, the multi-head self-attention weights of all its Decoder-Only modules are extracted, and after temporal causal masking constraints, multi-layer multi-head aggregation and spectral truncation denoising, the corresponding causal matrix is ​​generated.

[0085] The causal matrix corresponding to each time-series causal model is obtained by processing it according to the following steps:

[0086] First, the masked attention weights of all Decoder-Only modules and attention heads are aggregated to obtain the attention aggregation matrix:

[0087] ;

[0088] Attention aggregation matrix Perform feature decomposition and retain the first feature. Perform spectral truncation and reconstruction on each principal component to obtain the truncated attention matrix:

[0089] ;

[0090] The truncated attention matrix is ​​normalized in magnitude and then subjected to softmax mapping to obtain the final causal matrix:

[0091] ;

[0092] in, For indexing time-series causal models; For the first Each time-series causal model corresponds to A causal matrix of size, the nth element in the matrix Line 1 Column elements are denoted as , Representative node Pointing to node The causal strength, The total number of ground support operation nodes for flights; Index for Decoder-Only modules; For attention head index; and These represent the number of Decoder-Only modules within the model and the number of heads for multi-head attention within a single module, respectively. For the first In the first time-series causal model, the... The first Decoder-Only module The original attention weight matrix of the head; The element-wise Hadamard product of the matrix; A temporal causal mask matrix with values ​​of 0 or 1 retains only directed causal relationships that conform to the sequential constraints of flight support, while blocking reverse and cross-temporal invalid causal connections. The attention aggregation matrix is ​​obtained by summing and averaging the masked attention weights of all Decoder-Only modules and all attention heads, with a size of [size missing]. ; For attention aggregation matrix After performing eigenvalue decomposition, the front An orthogonal matrix formed by the eigenvectors corresponding to the largest eigenvalues; For the front eigenvalues ​​of order 1; To truncate the rank of the spectrum, only the first rank is retained. One principal component is used to filter noise; The maximum value among all eigenvalues ​​is used for overall matrix magnitude normalization.

[0093] Specifically, the temporal causal mask matrix For size A matrix of 0s and 1s, where the row index corresponds to the cause node. Column index corresponds to the result node The values ​​are strictly determined in accordance with the directed acyclic process topology of flight support.

[0094] The causal matrix construction method of the present invention ensures the effectiveness of causal inference through a two-level noise reduction constraint mechanism, and its specific beneficial effects are as follows:

[0095] First, the head-level gating mask ensures the physical consistency of causal direction. A 0-1 temporal causal mask with matching flight schedules to ensure directed acyclic topology is used to constrain the attention heads of each Decoder-Only module one by one, blocking invalid associations with reverse causality and no topological reachability, while suppressing the contribution of noise heads and improving the signal-to-noise ratio of the aggregated attention matrix.

[0096] Secondly, spectral truncation improves the robustness of the causal matrix. Eigenvalue decomposition is performed on the aggregated attention matrix, retaining the first eigenvalues. The principal components complete the low-rank reconstruction, filter out low-amplitude noise and weak random associations, reduce the disturbance of training data fluctuations on causal strength estimation, and ensure the stability of the core transmission structure.

[0097] The synergistic effect of these mechanisms can provide reliable causal quantitative support for subsequent delay risk scoring and intelligent scheduling decisions.

[0098] S3. Configure the delay root cause node information module using the cluster center vectors, the trained time-series causal model and the corresponding causal matrix, and construct the scheduling strategy dataset based on historical operation and maintenance data and the delay root cause node information module.

[0099] The Delay Root Cause Information module is used to generate delay root cause information for the input data.

[0100] The scheduling policy dataset in step S3 includes the historical scheduling policy dataset and the strong model distillation dataset.

[0101] The specific process of constructing the historical scheduling strategy dataset is as follows: Obtain scheduling decision data and corresponding historical operation and maintenance data; use the historical operation and maintenance data as the data to be inferred; input the data to be inferred into the delay root cause node information module to output the delay root cause node information of each job node; use the historical operation and maintenance data and the delay root cause node information of each job node as sample input; use the historical scheduling decision data corresponding to the historical operation and maintenance data as sample output labels to construct a historical scheduling strategy data sample.

[0102] The specific process of constructing the strong model distillation dataset is as follows: Obtain historical operation and maintenance data for which no corresponding scheduling decision data exists. Use this historical operation and maintenance data as the data to be inferred. Input this data into the delay root cause node information module to output the delay root cause node information for each job node. Input the historical operation and maintenance data and the delay root cause node information for each job node into the strong model to output the scheduling decision results. Apply a phased preference labeling strategy to the output scheduling decision results, retaining the preferred labels as accepted scheduling decision results. Use the retained scheduling decision results as sample output labels, and use the corresponding historical operation and maintenance data and the delay root cause node information for each job node as sample inputs to construct the strong model distillation data samples.

[0103] To verify the beneficial effects of this invention in creating a strong model distillation dataset, the following experiment was conducted in this embodiment: Only historical operation and maintenance data was input into the strong model to output scheduling decision results, and airport scheduling experts performed a first-stage manual preference labeling. The results showed that the acceptance rate of this manual preference labeling was significantly lower than that proposed in this invention, which inputs both historical operation and maintenance data and the root cause information of delays for each operational node into the strong model. This demonstrates the necessity of delay root cause information for generating high-quality scheduling strategies.

[0104] Specifically, since the historical scheduling strategy dataset is limited in actual processes, it is necessary to construct a strong model distillation dataset.

[0105] In this embodiment, the historical scheduling strategy dataset accounts for only 20% of the scheduling strategy dataset. By using a strong model distillation method, the original 20% of the scheduling decision data is expanded to cover all samples, which greatly improves the quality and sample space of the training dataset.

[0106] In practice, when performing DPO reinforcement training on the scheduling strategy model, each data sample in the scheduling strategy dataset is as follows:

[0107] The historical scheduling strategy dataset is shown in Table 2 below:

[0108] Table 2:

[0109]

[0110] The distillation dataset for the strong model is shown in Table 3 below:

[0111] Table 3:

[0112]

[0113] The phased preference labeling strategy consists of two phases. In the first phase, airport scheduling experts manually label the scheduling decision results output by the strong model as either accepting or rejecting them. In the second phase, a preference model is used to label the scheduling decision results output by the strong model.

[0114] The preference model is trained by the following steps: the scheduling decision results of the first stage are used as input and the manual preference labeling is used as output to construct seed samples. The preference model is trained based on the seed samples to obtain the preference model used for the preference labeling in the second stage.

[0115] In this embodiment, the strong model uses DeepSeek-v3.1, the preference model uses Qwen3-8B, and the negative sample generation model uses Qwen3-0.6B.

[0116] The specific processing flow of the delay root cause node information module is as follows:

[0117] D1. Match the data to be inferred with each cluster center vector to obtain the most matching cluster center vector. Input the data to be inferred into the time-series causal model corresponding to the most matching cluster center vector for processing to obtain the causal matrix and the delay time of each task node.

[0118] D2. The delay amplification factor is calculated by combining the delay time of each operation node with the topological path in the process and the scale of downstream reachable nodes of each node. Based on the preset propagation scale parameters and Weibull shape parameters of each node, and combined with the risk recursion relationship of downstream immediate process nodes, the propagation risk intensity of each operation node is calculated by the fixed point iteration method.

[0119] D3. Calculate the delay risk score for each job node based on the causal matrix of the corresponding trained time-series causal model, the node propagation risk intensity of each job node, and the corresponding delay amplification factor.

[0120] D4. Summarize the propagation risk intensity, delay amplification factor, and delay risk score of each node to generate the delay root cause node information required for the scheduling model input.

[0121] The delay risk score for each task node is set according to the following formula:

[0122] ;

[0123] ;

[0124] ;

[0125] in, , and All are indexes; For nodes Delay risk score; This represents the total number of task nodes. For the first The nodes in the causal matrix corresponding to each time-series causal model For nodes The causal strength value (i.e., the first) Line 1 Column values, nodes As the cause node, node For the result node, the causal direction is... (This aligns with the upstream and downstream propagation direction of the assurance process). It is a nonlinear exponent of causal strength; For nodes With nodes The number of edges in the shortest directed path in the guarantee process (if nodes) Unable to reach the node The number of edges in the shortest directed path distance is denoted as infinity. The topological distance Gaussian decay width parameter; For nodes The intensity of the risk of transmission; For propagation-amplification coupling coefficient; For nodes The delay amplification factor; and They are nodes and nodes The delay time; For nodes The propagation scale parameter reflects the node's tolerance for delays. The larger the parameter value, the more sufficient the node's delay buffer space, and the longer the delay duration is required to trigger a significant downstream propagation risk. is the gamma function, used to achieve scale normalization of the Weibull distribution; For nodes The Weibull shape parameter is used to characterize the nonlinear growth characteristics of the delay propagation risk at this node; The cascade coupling coefficient; For nodes The intensity of the risk of transmission; For nodes The set of downstream adjacent nodes, that is, ensuring that the nodes in the flowchart are connected. The next process node is directly connected by a directed edge; The causal strength coupling index; It is also a node index; and They are slave nodes and nodes The set of all directed paths to the final node (the final node is the endpoint of the guarantee process, i.e., the takeoff node). To ensure the nodes in the process The total number of downstream reachable nodes, i.e., from node The number of all downstream nodes that can be reached from the starting point via any directed path is used to measure the scale of the node's influence in the entire process. It is a directed path; For directed paths The edges in; For directed paths The number of sides; The path length decay exponent; For the first In the causal matrix corresponding to each time-series causal model, the edges The causal strength value.

[0126] Specifically, The essence is and Equivalent if there is a directed edge The starting point is the node The endpoint is node ,but Similarly, values ​​are retrieved from the causal matrix.

[0127] Intensity of transmission risk As defined recursively, a fixed-point iterative method is used for solution: during initialization, only the node's own delay is considered to obtain the initial value. Then iterates and updates round by round until the results of two adjacent rounds are updated. The solution is complete when the difference is less than the preset convergence threshold.

[0128] Flight ground support processes are not strictly linear and sequential execution structures, but rather present a directed acyclic network topology: core passenger procedures have fixed sequential constraints, while some ground support procedures can be carried out in parallel. Therefore, there are multiple different directed operation paths from the starting node of docking to the final node of takeoff.

[0129] The complete operational process includes: docking at the aircraft bridge, passenger disembarkation, baggage and cargo unloading, aircraft inspection, cleaning, baggage and cargo loading, passenger boarding, pushback, and takeoff. The sequential logic of each step is as follows: docking at the aircraft bridge is the starting point of the process. After the aircraft is docked at the bridge, the three independent steps of passenger disembarkation, baggage and cargo unloading, and aircraft inspection can be started simultaneously. Cabin cleaning can only begin after passenger disembarkation is completed. Baggage and cargo loading can only begin after baggage and cargo loading, aircraft inspection, and cleaning are all completed. Passenger boarding can only begin after passenger boarding is completed, followed by pushback and takeoff. Takeoff is the final step of the process.

[0130] Based on the above process logic, there are multiple differentiated operational paths from docking at the bridge to takeoff, for example:

[0131] Route 1 (Main Cabin Operations): Dock at bridge → Passengers disembark → Cleaning operations → Passengers board → Pushback → Takeoff.

[0132] Route 2 (Main Baggage Handling Line): Dock at the bridge → Baggage and cargo unloading → Baggage and cargo loading → Passenger boarding → Pushback → Takeoff.

[0133] Route 3 (Main Maintenance Inspection): Dock to bridge → Walk around the aircraft for inspection → Passenger boarding → Pushback → Takeoff.

[0134] The above-mentioned set of all directed paths from any job node to the final node corresponds to the set of directed paths to that node in the formula. This provides a topological basis for calculating path-based delay propagation intensity and solving for delay amplification factors. The structure of multiple parallel branches also results in significant differences in delay propagation paths and impact ranges at different nodes, demonstrating the technical necessity of full-path-dimensional risk quantification.

[0135] In this embodiment, the causal strength nonlinear index Topological distance Gaussian decay width parameter Propagation-Amplification Coupling Coefficient Total number of work nodes Cascade coupling coefficient Causal coupling strength index Path length decay index .

[0136] In this embodiment, considering the process characteristics, buffer space, and delay propagation sensitivity of the nine work nodes, this embodiment adopts three differentiated configurations. and One-to-one correspondence:

[0137] 1. High-tolerance nodes (sufficient buffer, delays are not easily propagated): These processes have high scheduling flexibility, and minor delays can be recovered through subsequent parallel operations and resource allocation. They correspond to large-scale parameters and small-shape parameters, therefore the following settings are made:

[0138] Bridge node: min, Inspection points around the machine: min, .

[0139] 2. Medium tolerance nodes (fixed workload, stable transmission characteristics): The workload of this type of process is relatively fixed, with some operational flexibility. However, once the delay exceeds the threshold, it will stably propagate downstream. Corresponding to medium parameter configuration, the settings are as follows:

[0140] Baggage and cargo unloading nodes: min, Cleaning operation nodes: min, Baggage and cargo loading nodes: min, .

[0141] 3. Low-tolerance nodes (strong constraints, rapid propagation of delays): These processes are heavily constrained by resource quotas and workflow connections, leaving almost no buffer space for delays. Even slight timeouts will quickly spread downstream. They correspond to small-scale parameters and large-shape parameters, therefore the following settings are provided:

[0142] Passenger disembarkation point: min, Passenger boarding points: min, Launch node: min, Takeoff node: min, .

[0143] The delay risk scoring formula set in this embodiment is based on the directed acyclic process topology of flight ground support, and is constructed by integrating causal strength, nonlinear characteristics of delay propagation and cascading amplification effect. The global impact is quantified by the causal strength of all nodes weighted by topological distance, the self-propagation risk is characterized by the node-specific Weibull kernel density combined with downstream recursive cascading, and then the amplification factor composed of the self-delay ratio, the full path propagation capability and the scale of downstream impact is superimposed. The delay risk of each operation node is quantified layer by layer from three dimensions, which is in line with the actual law of delay propagation in airport support.

[0144] This formula breaks through the limitations of traditional single-point delay duration statistics, and can accurately quantify the actual delay impact range and risk level of nodes. The use of node-differentiated parameters can adapt to the propagation characteristics of different processes, and at the same time fully adapt to the network topology structure with multiple parallel branches and multiple operation paths. Moreover, the parameters can be fitted with historical data and the solution method is mature and stable, with strong engineering applicability. It can provide quantitative decision support for flight delay early warning, dynamic resource allocation and key node control.

[0145] S4. Initialize the scheduling strategy model. Train the model using DPO-based reinforcement learning with the constructed scheduling strategy dataset to obtain the completed reinforcement training scheduling strategy model and complete local deployment.

[0146] The total loss function for reinforcement training of the scheduling strategy model is set according to the following formula:

[0147] ;

[0148] ;

[0149] ;

[0150] ;

[0151] in, These are the network parameters to be trained for the scheduling strategy model; The total loss function for training the scheduling strategy model; Standard DPO loss; To encourage the exploration of the dynamic weight coefficients of the entropy loss term, the weights are dynamically adjusted with the number of training iterations and the state of the loss. This is the current training round; Total number of training rounds; This represents the difference in total loss between two adjacent training rounds; This is the initial value of the entropy adjustment coefficient; This represents the weight decay coefficient for the step dimension; This is the adjustment factor for the dimension of loss change; This is the sensitivity coefficient to changes in loss; To encourage the exploration of entropy loss; This is the entropy adjustment coefficient that is dynamically updated as the training process progresses. This is the set of feasible scheduling strategies under the current scheduling state, i.e., given the information on the root cause nodes of the delay. At that time, it is the set of all legal scheduling schemes that satisfy all hard operational constraints of flight ground support; Information on the root causes of delays is input into the model; Output scheduling decision results for the scheduling strategy model; Input delay root cause node information In this case, the scheduling strategy model outputs the scheduling decision result. The probability value; It is a natural constant.

[0152] Specifically, hard operational constraints include three categories: resource capacity constraints, process sequence constraints, and civil aviation safety compliance constraints. Only the scheduling decision results within the set participate in the summation calculation of entropy regularization loss. Infeasible solutions that violate the constraints have a weight of 0 and are not included in the loss calculation.

[0153] The total loss function set in this invention uses the standard DPO loss as the main body to align the scheduling strategy with business preferences, and additionally adds an entropy loss to encourage exploration to improve the strategy exploration capability during the training phase. The weight of the regularization term is dynamically adjusted by coupling the training rounds and the loss changes of adjacent steps. At the same time, it relies on the entropy adjustment coefficient updated with the training process to change the form of entropy constraints. The entropy loss is calculated only for feasible scheduling actions that meet the time, resource, and security constraints. The solution is completed entirely based on the model's own output probability, without the introduction of any external fixed prior distribution.

[0154] The total loss function set in this invention removes the external reference distribution that is difficult to calibrate and cannot participate in gradient updates. All calculation terms can be derived by backpropagation of the model parameters, making the engineering implementation process simple and unambiguous. The dual-dimensional dynamic weights and variable entropy coefficients can adaptively balance the needs of global exploration in the early stage of training and stable convergence in the later stage of training. This avoids the strategy from getting trapped in local optima in advance and ensures that the final output scheduling scheme fits the actual preferences of airport operation and maintenance. The model training convergence speed and the practicality of the scheduling scheme are improved simultaneously.

[0155] In this embodiment, the initial value of the entropy adjustment coefficient is... Steps dimension weight decay coefficient Loss change dimension adjustment magnitude coefficient Loss change sensitivity coefficient Total training rounds Entropy adjustment coefficient The decays linearly with each training round, especially in the initial training phase. The linear decrease at the end of training period was reduced to At this point, we encourage the exploration of entropy loss degenerating into the standard Shannon entropy form.

[0156] Dynamic weights in the early stage of training The larger the value, the more it encourages exploration of the proportion of entropy loss and expands the search range of feasible scheduling schemes; as the number of training rounds increases, the weight gradually decays, and the total loss in the later stage of training is mainly the standard DPO loss, ensuring that the scheduling strategy fits the operational business preferences.

[0157] S5. Intelligent scheduling decisions are made based on real-time airport operation and maintenance data through the delay root cause node information module and the scheduling strategy model completed by reinforcement training.

[0158] Step S5 specifically involves: acquiring real-time airport operation and maintenance data, inputting the real-time operation and maintenance data into the delay root cause node information module for processing to obtain delay root cause node information, and inputting the real-time operation and maintenance data and delay root cause node information together into the scheduling strategy model that has undergone reinforcement training for inference and outputting scheduling decision results.

[0159] In this embodiment, the scheduling strategy model specifically adopts Qwen3-8B.

[0160] Thus, from steps S1 to S5, this invention constructs a real-time closed loop of "node-level causal prediction - accurate root cause scoring - intelligent dynamic decision-making" through the delay root cause node information module and scheduling strategy model, and uses an improved DPO loss function for reinforcement training. Under the premise of strictly adhering to physical constraints, it generates diverse and adaptive scheduling decision results, significantly reduces the scope of delay propagation, optimizes resource allocation efficiency, and realizes the paradigm shift of airport operation and maintenance from passive response to proactive intervention.

[0161] To verify the beneficial effects of the novel improved loss function used in the reinforcement training of the scheduling strategy model in this invention, the following ablation experiments were conducted in this embodiment, as shown in Table 4 below:

[0162] Table 4:

[0163]

[0164] The evaluation indicators set in this embodiment are shown in Table 5 below:

[0165] Table 5:

[0166]

[0167] Specifically, the decision repetition rate measures the diversity of scheduling decision results generated by the scheduling strategy model under the same input scenario. In actual airport scheduling scenarios, there is often more than one reasonable scheduling strategy for the same delay situation. If the model repeatedly generates semantically highly similar or even identical scheduling decision results when faced with the same input multiple times, it indicates that the model has a decision pattern collapse problem, lacks the ability to explore diverse scheduling schemes, and cannot provide scheduling personnel with a rich set of alternatives.

[0168] Root cause hit rate is used to measure whether the scheduling decision results generated by the scheduling strategy model can effectively respond to the key bottleneck nodes identified by the delay root cause node information module. This invention, through the delay root cause node information module, outputs a delay risk score for each job node in each scheduling scenario. The node with the highest score (i.e., the Top-1 root cause node) is identified as the bottleneck link that most needs priority scheduling in the current scenario. If the scheduling strategy model can make targeted resource scheduling or process adjustments for the Top-1 root cause node in its generated scheduling decision results, it indicates that the model's decision is highly aligned with the delay root cause analysis results.

[0169] The results of the ablation experiment are shown in Table 6 below:

[0170] Table 6:

[0171]

[0172] This invention, through adaptive encouragement of exploration entropy loss, reduces the decision repetition rate to 8.27% while maintaining a high root cause hit rate (>94%), achieving the triple objectives of "low violation, high diversity, and high accuracy." Compared to the scheme that only introduces a single loss (Experiment B), this invention achieves optimal results on both metrics, demonstrating the effectiveness of the synergistic effect of physical constraints and exploration encouragement, and also proving the importance of root cause information input for the scheduling strategy of large models.

[0173] This invention constructs a closed-loop method for the entire chain of "node-level causal prediction - root cause accurate scoring - intelligent dynamic decision-making". It achieves the organic integration of operation node-level delay prediction and causal reasoning through a time-series causal model, accurately identifies key root cause nodes and their propagation effects by combining a multi-dimensional scoring model, and generates a multi-dimensional scheduling strategy that strictly adheres to physical constraints by using an improved DPO reinforcement learning framework. This forms a closed-loop mechanism from real-time perception to intelligent decision-making, significantly reducing the scope of delay propagation, optimizing resource allocation efficiency, and realizing a paradigm shift in airport operation and maintenance from passive response to proactive intervention.

[0174] The above embodiments are merely preferred embodiments provided to fully illustrate the present invention, and the scope of protection of the present invention is not limited thereto. Equivalent substitutions or modifications made by those skilled in the art based on the present invention are all within the scope of protection of the present invention. The scope of protection of the present invention is defined by the claims.

Claims

1. A method for temporal causal reasoning and intelligent processing of flight operation delays for smart airports, characterized in that, Includes the following steps: S1. Obtain historical airport operation and maintenance data and perform cleaning, feature processing and clustering processing in sequence to obtain several data subsets and each cluster center vector; S2. Construct a time-series causal model. Train the corresponding time-series causal model for each data subset to obtain several trained time-series causal models, and extract the causal matrix corresponding to each time-series causal model. S3. Configure the delay root cause node information module using each cluster center vector, the trained time-series causal model and the corresponding causal matrix, and construct the scheduling strategy dataset based on historical operation and maintenance data and the delay root cause node information module. The specific processing flow of the delay root cause node information module is as follows: D1. Match the data to be inferred with each cluster center vector to obtain the most matching cluster center vector. Input the data to be inferred into the time-series causal model corresponding to the most matching cluster center vector for processing to obtain the causal matrix and the delay time of each task node. D2. The delay amplification factor is calculated by combining the delay time of each operation node with the topological path in the guarantee process and the scale of downstream reachable nodes of each node. Based on the preset propagation scale parameters and Weibull shape parameters of each node, and combined with the risk recursion relationship of the downstream subsequent process nodes, the fixed-point iteration method is used to calculate the propagation risk intensity of each operation node. D3. Calculate the delay risk score for each work node based on the causal matrix, the node propagation risk intensity of each work node, and the corresponding delay amplification factor. D4. Summarize the propagation risk intensity, delay amplification factor, and delay risk score of each node to generate the delay root cause node information required for the input of the scheduling strategy model; S4. Initialize the scheduling strategy model, and train the model using DPO-based reinforcement learning through the constructed scheduling strategy dataset to obtain the completed reinforcement training scheduling strategy model and complete local deployment. S5. Intelligent scheduling decisions are made based on real-time airport operation and maintenance data through the delay root cause node information module and the scheduling strategy model completed by reinforcement training.

2. The method for causal reasoning and intelligent processing of flight delays according to claim 1, characterized in that, Step S1 specifically involves: S11. Obtain historical airport operation and maintenance data and perform cleaning and feature processing in sequence to obtain processed historical operation and maintenance data. The historical operation and maintenance data includes ground support data and scheduling decision data for each flight in the airport for a continuous number of historical periods. The flight ground support data includes data from multiple operational nodes; S12. Extract the delay time corresponding to each operation node from the processed flight ground support data. Use the flight ground support data as the sample input and the delay time of each operation node as the sample output label to construct a single data sample and obtain the delay root cause dataset. S13. Use K-Means to cluster the delay root cause dataset to obtain several data subsets and the cluster center vector corresponding to each data subset.

3. The method for causal reasoning and intelligent processing of flight delays according to claim 2, characterized in that: The temporal causal model in step S2 includes an embedding coding layer, a backbone network, and a prediction head. Flight ground support data in the data subset is input into the embedding coding layer for processing to obtain multi-scale features. The multi-scale features are input into the backbone network, which includes multiple cascaded Decoder-Only modules. The last Decoder-Only module outputs multi-dimensional features. The multi-dimensional features are then processed by the prediction head to output the delay time of each operation node in the flight ground support data. The specific processing flow of the embedded coding layer is as follows: the flight ground support data in the data subset is first processed by three parallel one-dimensional convolutions to obtain different feature maps, and the three feature maps are then concatenated to obtain the transformation fusion feature; The input data is processed using RoPE rotational position coding to obtain positional coding features. Time intervals in the input data are encoded to obtain temporal coding features. The positional coding features and temporal coding features are then gated and fused to obtain spatiotemporal coding features. The spatiotemporal coding features are then added to the transform fusion features to obtain multi-scale features. The specific processing flow of each Decoder-Only module is as follows: the module input is first normalized by a layer and then fed into a multi-head mask self-attention module to obtain self-attention features. The self-attention features and the module input residuals are then concatenated to obtain preliminary fused features; The fused features are then passed through layer normalization and feedforward layers to obtain feedforward features. The feedforward features and module input are then residually connected and used as the output of the Decoder-Only module. For each trained temporal causal model, the weights of all attention points in the model are taken to construct the corresponding causal matrix.

4. The method for causal reasoning and intelligent processing of flight delays according to claim 1, characterized in that: The scheduling strategy dataset in step S3 includes a historical scheduling strategy dataset and a strong model distillation dataset. The specific process of constructing the historical scheduling strategy dataset is as follows: acquire scheduling decision data and corresponding historical operation and maintenance data, use the historical operation and maintenance data as the data to be inferred, input the data to be inferred into the delay root cause node information module to output the delay root cause node information of each job node; use the historical operation and maintenance data and the delay root cause node information of each job node as sample input, use the historical scheduling decision data as sample output labels, and construct historical scheduling strategy data samples. The specific process of constructing the strong model distillation dataset is as follows: acquire historical operation and maintenance data that does not have corresponding scheduling decision data, use the historical operation and maintenance data as the data to be inferred, input the data to be inferred into the delay root cause node information module to output the delay root cause node information of each job node; input the historical operation and maintenance data and the delay root cause node information of each job node together into the strong model to output the scheduling decision result, use a phased preference labeling strategy to perform preference labeling on the output scheduling decision result, and retain the preference label as the accepted scheduling decision result. The retained scheduling decision results are used as the sample output labels, and the corresponding historical operation and maintenance data and the delay root cause node information of each job node are used as the sample inputs, thereby constructing a strong model distillation data sample.

5. The flight operation delay temporal causal reasoning and intelligent processing method according to claim 4, characterized in that: The phased preference labeling strategy includes two phases. In the first phase, airport scheduling experts manually label the scheduling decision results output by the strong model as either accepting or rejecting them. In the second phase, a preference model is used to label the scheduling decision results output by the strong model. The preference model is trained by the following steps: taking the scheduling decision results of the first stage as input and the manual preference labeling as output to construct seed samples, training the preference model based on the seed samples, and thus obtaining the preference model for preference labeling in the second stage.

6. The method for causal reasoning and intelligent processing of flight delays according to claim 1, characterized in that: The delay risk score for each task node in step D3 is set according to the following formula: ; ; ; in, , All are node indexes; For indexing time-series causal models; Assess the risk of delays; This represents the total number of task nodes. For the first Nodes in a causal matrix For nodes The causal strength value; It is a nonlinear exponent of causal strength; Let be the number of edges in the shortest directed path between the two nodes in the guarantee process; The topological distance Gaussian decay width parameter; To assess the intensity of the risk of transmission; For propagation-amplification coupling coefficient; This is the delay amplification factor; and All of these are delay times; For propagation scale parameters; It is the gamma function; For Weibull shape parameters; The cascade coupling coefficient; To assess the intensity of the risk of transmission; For the set of downstream adjacent nodes; The causal strength coupling index; For node indexing; and They are slave nodes and nodes The set of all directed paths to the final node; This represents the total number of downstream nodes. It is a directed path; For directed paths The edges in; For directed paths The number of sides; The path length decay exponent; For the first In the causal matrix, the edges The causal strength value.

7. The method for causal reasoning and intelligent processing of flight delays according to claim 1, characterized in that: The loss function used for reinforcement training of the scheduling strategy model is set according to the following formula: ; ; ; ; in, These are the parameters of the network to be trained; This is the total loss function; For DPO loss; To encourage exploration of dynamic weighting coefficients for the entropy loss term; This is the current training round; Total number of training rounds; This represents the difference in total loss between two consecutive training rounds. This is the initial value of the entropy adjustment coefficient; This represents the weight decay coefficient for the step dimension; This is the adjustment factor for the dimension of loss change; This is the sensitivity coefficient to changes in loss; To encourage the exploration of entropy loss; This is the entropy adjustment coefficient that is dynamically updated as the training process progresses. This represents the set of feasible scheduling strategies under the current scheduling state. The information on the root causes of delays is input into the model; To output the scheduling decision results; Input delay root cause node information In this case, the scheduling strategy model outputs the scheduling decision result. The probability value; It is a natural constant.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 7.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 7.

Citation Information

Patent Citations

  • Flight delay prediction method based on deep learning

    CN110659773A

  • Flight delay prediction method based on multi-modal data fusion and large model fine tuning

    CN120633951A

  • Airport operation guarantee capability evaluation and analysis method and system

    CN121212904A

  • Flight delay prediction method and system based on sweep evolution

    CN117391236A

  • Flight ground support event time constraint optimization method based on causal reasoning

    CN122175019A