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

Abstract

This invention provides a time-constrained optimization method for flight ground support events based on causal reasoning, comprising the following steps: Step 1, construction of a hierarchical analysis system and data standardization processing; Step 2, generation of civil aviation prior constraint set and causal feasible region matrix; Step 3, causal structure learning using the constraint-based NOTEARS-MCP algorithm; Step 4, partitioning of heterogeneous intervals of influencing factors and generation of effective intervention sets; Step 5, quantification and intensity ranking of entropy balance weighted causal effects; Step 6, Markov decision process modeling and space compression of the support process; Step 7, construction of a causal dual-constraint initial behavioral strategy set; Step 8, solving for the global optimal strategy using the path integral control algorithm; Step 9, robustness verification of the optimal strategy under perturbation; Step 10, output and mapping of time control requirements for flight ground support events. This invention effectively reduces the risk of flight departure delays and helps airports achieve refined and intelligent management of support processes.

Description

Technical Field

[0001] This invention relates to the field of civil aviation airport operation management and flight ground support optimization technology, and in particular to a time constraint optimization method for flight ground support events based on causal reasoning. Background Technology

[0002] Flight departure delays have become a core pain point hindering the improvement of service quality in the civil aviation industry. Industry data shows that departure delays caused by abnormalities in ground handling account for a consistently high proportion of total delays. Ground handling activities are the only core aspect that airport operations management can proactively intervene in and directly control. Therefore, developing scientific and reasonable time constraint strategies for handling events and improving the efficiency of the entire handling process are key issues in alleviating flight departure delays.

[0003] Existing technical solutions for optimizing flight ground support mainly suffer from three core defects: Firstly, there are support process optimization schemes based on simulation modeling, system dynamics, and discrete event analysis. These schemes mostly focus on buffer time setting, ground equipment resource allocation, and support activity dependency identification. The core of these schemes is based on the correlation analysis between variables, which cannot characterize the real intervention effect of different influencing factors on the efficiency of support event execution. It is difficult to identify the core driving reasons for changes in support efficiency. They can only achieve post-event analysis and static simulation, and cannot provide a feasible quantitative decision-making basis for support management in dynamic operation scenarios. Secondly, there are probabilistic decision-making schemes based on historical statistics. These schemes construct decision-making strategies by analyzing the statistical frequency of variable co-occurrence in historical data. The implicit assumption is that the observed distribution can fully represent the real decision-making environment. In flight support scenarios with a large number of confounding factors, it is easy to learn spurious correlations between variables rather than real causal relationships. The performance of the strategy is severely affected by data bias. Not only does the decision-making logic lack interpretability, but its generalization ability, anti-interference ability, and robustness in complex operating environments are also significantly insufficient. Third, existing research on introducing causal modeling and reinforcement learning into the field of civil aviation operations has significant technical limitations. Causal modeling solutions primarily focus on discovering causal relationships between nodes in a support event, failing to conduct quantitative analysis of the causal effects of specific influencing factors, and neglecting to deeply integrate causal information with decision optimization, thus failing to formulate time-constrained strategies that can directly guide support execution. Reinforcement learning-related optimization solutions, lacking the constraints and guidance of causal prior information, exhibit low search efficiency in the high-dimensional state-action space of flight support, easily converging to local optima, and failing to achieve global collaborative optimization across multiple support stages, making it difficult to balance time constraints across stages to achieve an overall improvement in flight departure punctuality.

[0004] In summary, existing technologies cannot accurately depict the true causal relationship between factors affecting flight ground support and support events, cannot quantify the real impact of different control behaviors on support efficiency, and are difficult to achieve globally optimal time-constrained decision-making throughout the entire support process. This has become a key technical bottleneck restricting the refined control of flight ground support and the improvement of flight departure punctuality. Summary of the Invention

[0005] This invention provides a time constraint optimization method for flight ground support events based on causal reasoning. By using causal reasoning technology to accurately characterize the potential causal structure between flight ground support events and influencing factors, it quantifies the real causal effects of different intervention behaviors on support events, and then deeply integrates causal information with a reinforcement learning decision framework to construct and solve the globally optimal time constraint strategy for the entire flight support process. This solves the problems of decision bias, poor interpretability, insufficient global optimization capability, and weak strategy robustness in existing technologies, effectively reducing the risk of flight departure delays and helping airports achieve refined and intelligent management of support processes.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: The time-constrained optimization method for flight ground support events based on causal reasoning includes the following steps: Step 1: Obtain the time node dataset of the entire historical flight support process, then construct a hierarchical analysis system for flight ground support, determine the key events of the target layer support and the corresponding control factors of the influencing layer, design a unified standardized quantitative reference quantity, and obtain the standardized quantitative reference quantity dataset after preprocessing. Step 2: Combining the hierarchical analysis system of flight ground support, construct a set of prior constraints for civil aviation business and generate a causal structure feasible domain matrix; Step 3: Based on the standardized quantitative reference dataset, and with the civil aviation business prior constraint set and the causal structure feasible region matrix as constraints, the NOTEARS-MCP algorithm with adjacency matrix smoothing constraint is used to learn the sparse causal adjacency matrix and causal graph of flight ground support events. Step 4: Based on the standardized quantitative reference dataset and the sparse causal adjacency matrix, the standardized quantitative reference quantities of the factors affecting flight ground support are divided into interval heterogeneity. After compliance verification, the effective intervention interval set corresponding to each key event is generated. Step 5: Based on the standardized quantitative reference dataset, causal graph, and effective intervention interval set, the causal effect estimation algorithm based on entropy balance weighting is used to calculate the average intervention effect corresponding to each effective intervention interval and generate a causal intensity ranking table of intervention intervals. Step 6: Based on the sparse causal adjacency matrix, the average intervention effect and the causal intensity ranking table of the intervention interval, construct the Markov decision process model for flight ground support, and obtain the compressed state-action space set after state-action space compression. Step 7: Based on the average intervention effect, the causal intensity ranking table of the intervention interval, and the compressed state-action space set, construct an initial behavioral strategy set with dual constraints of causal intensity. Step 8: Based on the sparse causal adjacency matrix, average intervention effect, Markov decision process model, compressed state-action space set, and initial behavior strategy set, the path integral control algorithm based on linearly solvable Markov decision process is used to solve for the global optimal time constraint strategy for the entire process of flight ground support. Step nine: Perform robust perturbation verification on the globally optimal time constraint strategy to obtain the final globally optimal time constraint strategy that passes the verification.

[0007] In this manual, the flight ground support event time constraint optimization method based on causal reasoning also includes step ten, which transforms the final global optimal time constraint strategy into a feasible time control requirement for the entire flight ground support process by combining the hierarchical analysis system and standardized quantitative reference calculation rules of flight ground support.

[0008] In this specification, the civil aviation business prior constraint set constructed in step two includes two types: prohibitive constraints and mandatory constraints. Prohibitive constraints are used to eliminate causal directions that do not conform to the business time sequence, while mandatory constraints are used to retain causal dependency edges that conform to the business logic. The causal structure feasible region matrix is ​​a binary matrix generated based on the two types of constraints, which is used to limit the search boundary of the causal structure.

[0009] In this specification, the NOTEARS-MCP algorithm with adjacency matrix smoothing constraint in step three transforms the discrete search of the directed acyclic graph into a continuous optimization problem. The objective function is constructed using the data fit goodness term, the MCP non-convex penalty term, the prior constraint term, and the acyclic smoothing constraint term. The objective function is then solved iteratively using the ADMM optimization algorithm, and finally, the sparse causal adjacency matrix is ​​obtained after thresholding.

[0010] In this manual, the interval heterogeneity partitioning in step four is first completed by using the equal frequency quantile partitioning method to complete the initial interval partitioning, and then the interval boundaries are adjusted in combination with the time scale of civil aviation ground support operations. After the compliance verification is completed by sample size verification and covariate overlap verification, the effective intervention interval set is generated after removing the intervals that do not meet the requirements.

[0011] In this specification, the causal effect estimation algorithm based on entropy balance weighting in step five aims to minimize the entropy loss of the control group sample weights and solves for the optimal weights under the constraint of complete balance of covariate moments between the intervention group and the control group. Based on the optimal weights, the mean difference in the execution time of the target key events between the intervention group and the weighted control group is calculated to obtain the average intervention effect of the corresponding intervention interval.

[0012] In this specification, in the Markov decision process model for flight ground support constructed in step six, state-action space compression is based on the causal intensity ranking table to eliminate invalid actions with low impact intensity, thereby reducing the strategy search space.

[0013] In this specification, the initial behavioral strategy set with causal strength dual constraints constructed in step seven is first calculated based on the average intervention effect to determine the initial causal weight of each action, and then the initial weight is corrected by the state change benefit of the action execution. After normalization, the selection probability of each action in each state is obtained, thus forming the initial behavioral strategy set.

[0014] In this specification, the path integral control algorithm based on a linearly solvable Markov decision process in step eight transforms the nonlinear Bellman dynamic programming equation into a linearly solvable equation through exponential transformation. It then iteratively solves the state value function using the initial behavior policy set to obtain the optimal action probability distribution and finally extracts the globally optimal time constraint policy for the entire process.

[0015] In this specification, the robust perturbation check in step nine applies a fixed proportion of positive and negative perturbations to the average intervention effect corresponding to all actions in the globally optimal time-constrained strategy. Based on the results after the perturbation, the causal strength ranking is recalculated, and the change rate of the ranking before and after the perturbation is compared to determine whether the strategy passes the check. If the check fails, it is fed back to step four to adjust the interval partitioning granularity, and the corresponding steps are repeated to solve for the optimal strategy again.

[0016] In summary, the present invention has at least the following beneficial effects: This invention effectively eliminates the interference of confounding factors through causal discovery and causal effect estimation. It can accurately identify the core driving factors affecting the efficiency of flight ground support, quantify the real impact of different control behaviors on support events, and provide interpretable quantitative basis with causal semantics for time control decisions in the support process. It fundamentally avoids the decision bias problem caused by spurious correlations.

[0017] This invention deeply integrates causal structure and causal effect information into the reinforcement learning decision-making process. On the one hand, it significantly reduces the search space of the policy through causal prior constraints, effectively accelerating the algorithm convergence process. On the other hand, it guides policy updates through quantified causal effect results, allowing the agent to focus on key behaviors that have a substantial driving effect on ensuring efficiency, thus significantly improving the accuracy of the decision-making policy and the optimization efficiency.

[0018] This invention completes the modeling of the entire flight ground support process based on Markov decision processes, realizing the leap from single-event local time constraint optimization to multi-stage global collaborative optimization, avoiding the global efficiency loss caused by the simple superposition of local optimal strategies in each stage, and can achieve the global optimal matching of time constraints for the entire support process.

[0019] The time constraint strategy generated by this invention has excellent robustness and generalization ability. It can effectively suppress the performance degradation of the strategy caused by fluctuations in the operating environment and data bias. It is adaptable to different operating scenarios with different loads and complexities at airports. It can stably support the refined management and control of airport ground support and effectively improve the on-time performance of flights and the level of airport operation management. Attached Figure Description

[0020] Figure 1 This is a flowchart illustrating the time constraint optimization method for flight ground support events based on causal reasoning involved in this invention.

[0021] Figure 2 This is a schematic diagram illustrating the process of data standardization and causal prior construction involved in this invention.

[0022] Figure 3 This is a schematic diagram of the process of causal structure learning and causal effect quantification involved in this invention.

[0023] Figure 4 This is a schematic diagram illustrating the process of solving the optimal strategy for causal enhancement involved in this invention. Detailed Implementation

[0024] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0025] like Figure 1 As shown, this embodiment provides a time constraint optimization method for flight ground support events based on causal reasoning, including: Step 1: Construction of a Hierarchical Analysis System and Standardization of Data: This step establishes a unified and traceable quantitative analysis foundation, primarily addressing industry pain points such as multiple ground support processes for civil aviation flights, inconsistent job evaluation dimensions, and inconsistent data definitions. The data standardization and causal prior construction process is as follows: Figure 2 As shown.

[0026] Obtain a dataset of the entire timeline of historical flight support. The basic data source is the original dataset of the entire timeline of historical flight support from domestic hub airports for two consecutive months. The dataset must include all traceable timestamps from flight arrival to departure, aircraft type information, gate information, and support personnel arrival time information, with a data integrity rate of no less than 98%.

[0027] A hierarchical analysis system for flight ground support is constructed, consisting of a target layer and an impact layer. From the entire flight ground support process, nine key ground support events with strict temporal relationships, independent controllability, and direct impact on flight departure punctuality are selected to form a fixed target layer set. A complete example is as follows: The first key ground support event is the placement of aircraft wheel chocks; the second is the docking of passenger boarding bridges; the third is the opening of cabin doors; the fourth is the disembarkation of arriving passengers; the fifth is cabin cleaning; the sixth is the boarding of departing passengers; the seventh is the closing of cabin doors; the eighth is the evacuation of passenger boarding bridges; and the ninth is the pushback and wheel chock removal.

[0028] Subsequently, the influencing factors of flight ground support were matched and fixed. For each key event in flight ground support, corresponding influencing factors that could be adjusted through pre-emptive control were matched, forming a one-to-one mapping relationship between the target layer and the influence layer. A complete example is as follows: The influencing factors for the key event of aircraft wheel chock placement are two: the arrival status of ground staff operating the wheel chocks at the gate and the aircraft's arrival status at the gate; the influencing factors for the key event of passenger boarding bridge docking are none; the influencing factors for the key event of cabin door opening are three: the punctuality of passenger boarding bridge docking, the arrival status of ground support personnel, and the efficiency of the passenger boarding bridge docking process; the influencing factors for the key event of arriving passengers disembarking are... The influencing factor for the following critical events is as follows: The arrival status of disembarkation support personnel is one factor; the influencing factors for cabin cleaning flights are three factors: the arrival status of cabin cleaning personnel, the disembarkation efficiency of arriving passengers, and the total cabin cleaning time; the influencing factors for departing passenger boarding flights are two factors: the arrival status of boarding service personnel and the boarding efficiency of departing passengers; the influencing factors for cabin door closing flights are two factors: the disembarkation efficiency of arriving passengers and the boarding efficiency of departing passengers; the influencing factor for passenger boarding bridge evacuation flights is one factor: the arrival status of boarding bridge operation ground staff at the aircraft stand; and the influencing factor for aircraft pushback and wheel chock removal flights is one factor: the towing vehicle docking efficiency.

[0029] Next, we completed the calculation of standardized quantitative reference quantities. For all key events and influencing factors in flight ground support, we designed a unified standardized quantitative reference quantity with time difference as the core. All reference quantities are calculated based on the input historical time node dataset. Time difference was chosen as the unified quantitative indicator because the time node data of civil aviation ground support is objective data, which can eliminate the differences in evaluation dimensions between different links and positions, and realize horizontal comparison of efficiency across the entire process.

[0030] The standardized quantitative reference formula for calculating the arrival time of ground staff operating wheel chocks at the aircraft position is as follows: In the formula, A standardized and quantified reference quantity for the arrival of ground crew at the machine position for wheel chock operation, in minutes; The timestamp of the flight's arrival at the gate; This is the timestamp for the arrival of ground crew at the machine position for wheel chock operation.

[0031] The standardized quantitative reference formula for calculating the aircraft's arrival status at the parking position is as follows: In the formula, A standardized quantitative reference quantity for the status of an aircraft arriving at its parking position, expressed in minutes; The timestamp indicating the completion of the aircraft's parking position check.

[0032] The standardized quantitative reference formula for calculating the on-time performance of passenger boarding bridge docking is as follows: In the formula, A standardized and quantifiable reference for the punctuality of passenger boarding bridge connections, expressed in minutes; The timestamp for when the cabin door was opened; The timestamp indicating the completion of the passenger boarding bridge docking.

[0033] The standardized quantitative reference formula for calculating the arrival status of ground support personnel is as follows: In the formula, A standardized and quantifiable reference quantity for the arrival status of ground support personnel, in minutes; The timestamp for the arrival of ground crew operating the cabin doors at the aircraft position.

[0034] The standardized quantitative reference formula for calculating the efficiency of the passenger boarding bridge docking process is as follows: In the formula, A standardized and quantifiable reference quantity for the efficiency of passenger boarding bridge docking process, in minutes; This is the timestamp of the boarding bridge operator arriving at the gate.

[0035] The standardized quantitative reference formula for calculating the arrival status of off-board support personnel is as follows: In the formula, A standardized and quantified reference quantity for the arrival status of off-board support personnel, in minutes; This is the timestamp of the off-board support personnel arriving at the aircraft position.

[0036] The standardized quantitative reference formula for calculating the arrival status of cabin cleaning personnel is as follows: In the formula, A standardized and quantified reference quantity for the arrival status of cabin cleaning personnel, in minutes; This is the timestamp of the cabin cleaning staff arriving at the aircraft position.

[0037] The standardized quantitative reference formula for calculating the arrival passenger disembarkation efficiency is as follows: In the formula, A standardized quantitative reference for passenger disembarkation efficiency, in minutes; A timestamp indicating the end of cabin cleaning; This is the timestamp indicating the end of the passenger's departure process.

[0038] The standardized quantitative reference formula for calculating the total cabin cleaning time is as follows: In the formula, A standardized quantitative reference for the total cabin cleaning time, in minutes; The timestamp indicating when cabin cleaning began.

[0039] The standardized quantitative reference formula for calculating the arrival status of boarding service personnel is as follows: In the formula, A standardized and quantifiable reference quantity for the arrival status of boarding service personnel, in minutes; The timestamp for when the boarding staff arrive at the gate.

[0040] The standardized quantitative reference formula for calculating departing passenger boarding efficiency is as follows: In the formula, A standardized and quantified reference quantity for departing passenger boarding efficiency, in minutes; This is the timestamp indicating the end of boarding for departing passengers. This is the timestamp indicating when departing passengers begin boarding.

[0041] The standardized quantitative reference formula for calculating the arrival status of boarding bridge ground staff at the aircraft stand is as follows: In the formula, A standardized and quantified reference quantity for the arrival status of ground staff operating boarding bridges at the aircraft stand, in minutes; The timestamp for the arrival of the evacuation personnel from the boarding bridge at the aircraft stand.

[0042] The standardized quantitative reference formula for calculating the tractor docking efficiency is as follows: In the formula, This is a standardized and quantified reference quantity for tractor docking efficiency, expressed in minutes. The timestamp indicating the completion of the docking between the tractor and the aircraft; This is the timestamp indicating the completion of passenger boarding bridge evacuation.

[0043] The standardized quantitative reference formula for calculating the critical event of aircraft wheel chock placement on ground support is as follows: In the formula, Standardized quantitative reference quantities for critical ground support events related to aircraft wheel chock placement, in minutes; This is the timestamp indicating when the wheel chocks were placed.

[0044] The standardized quantitative reference formula for calculating key events in ground support for passenger boarding bridge connections is as follows: In the formula, Standardized quantitative reference quantities for key events in passenger boarding bridge connection to flight ground support, in minutes.

[0045] The standardized quantitative reference formula for calculating the critical event of cabin door opening on flight ground support is as follows: In the formula, Standardized quantitative reference quantities for critical events in ground support for flight cabin door opening, in minutes; This is the timestamp indicating when the passenger began disembarking.

[0046] The standardized quantitative reference formula for critical events in ground support for arriving passengers disembarking flights is as follows: In the formula, This is a standardized, quantifiable reference quantity for critical ground support events for passengers disembarking from flights, measured in minutes.

[0047] The standardized quantitative reference formula for calculating critical events in cabin cleaning and ground support for flights is as follows: In the formula, Standardized quantitative reference quantities for critical events in cabin cleaning and ground support for flights, in minutes.

[0048] The standardized quantitative reference formula for critical ground support events for departing passenger boarding flights is as follows: In the formula, This is a standardized, quantified reference quantity for critical ground support events for departing passengers boarding flights, measured in minutes.

[0049] The standardized quantitative reference formula for calculating the critical event of cabin door closure on flight ground support is as follows: In the formula, Standardized quantitative reference quantities for critical events in ground support for flights with cabin door closure, in minutes; This is the timestamp indicating when the cabin door has closed.

[0050] The standardized quantitative reference formula for calculating critical events in ground support for passenger boarding bridge evacuation flights is as follows: In the formula, Standardized quantitative reference quantities for critical events in ground support for passenger boarding bridge evacuation of flights, in minutes.

[0051] The standardized quantitative reference formula for critical ground support events related to aircraft pushback and wheel chock removal is as follows: In the formula, Standardized quantitative reference quantities for critical ground support events related to aircraft rollout and wheel chock removal, in minutes; The timestamp indicating the removal of the wheel chocks.

[0052] Finally, the dataset was cleaned and standardized. Outlier truncation was performed on all the calculated standardized quantitative reference quantities, and outlier samples with values ​​exceeding 3 times the standard deviation were removed, thus forming a complete standardized quantitative reference quantity dataset.

[0053] Step 2, generation of civil aviation prior constraint set and causal feasible region matrix: The core function of this step is to define the search feasible region that conforms to the business logic of civil aviation, and to avoid the causal discovery algorithm from learning pseudo-causal relationships that are out of order or do not conform to the operating rules. It is the core link for adapting the general causal discovery method to the professional scenario of civil aviation.

[0054] The constraints in this step are derived from the currently effective Civil Airport Flight Ground Support Service Specifications, the Standard Operation Manual for Domestic Hub Airports, and the fixed time sequence logic of the entire flight ground support process. At the same time, the constraints are refined by combining the fixed mapping relationship set between the target layer and the influence layer generated in step one.

[0055] First, the prior constraint set for civil aviation operations is constructed. For the fixed temporal logic of the entire flight ground support process, two types of insurmountable hard constraints are constructed. A complete example is as follows: The first type is prohibitive constraints, used to prohibit causal directions that do not conform to the business temporal sequence. Specifically, this includes prohibiting directed causal edges from the passenger boarding bridge docking critical event to the aircraft wheel chock placement critical event; prohibiting directed causal edges from the cabin door opening critical event to the passenger boarding bridge docking critical event; and prohibiting directed causal edges from the arriving passenger disembarkation critical event to the cabin door opening critical event. Directed causal edges for ground support critical events are prohibited, specifically those from: a cabin cleaning flight ground support critical event to an arriving passenger disembarkation flight ground support critical event; a departing passenger boarding flight ground support critical event to a cabin cleaning flight ground support critical event; a cabin door closing flight ground support critical event to a departing passenger boarding flight ground support critical event; a passenger bridge evacuation flight ground support critical event to a cabin door closing flight ground support critical event; and a flight pushback and wheel chock removal flight ground support critical event. The first category is directed causal edges from the key event of aircraft wheel chock placement to the key event of passenger boarding bridge docking; the second category is mandatory constraints, used to force the retention of causal dependency edges that conform to business logic. Specifically, this includes forcibly retaining directed causal edges from the key event of aircraft wheel chock placement to the key event of passenger boarding bridge docking; forcibly retaining directed causal edges from the key event of passenger boarding bridge docking to the key event of cabin door opening; forcibly retaining directed causal edges from the key event of cabin door opening to the key event of arriving passenger disembarkation; and forcibly retaining directed causal edges from the key event of arriving passenger disembarkation to the key event of arriving passenger disembarkation. The directed causal edges from critical ground support events to critical ground support events for cabin cleaning flights are forcibly retained. The directed causal edges from critical ground support events for cabin cleaning flights to critical ground support events for departing passenger boarding flights are forcibly retained. The directed causal edges from critical ground support events for departing passenger boarding flights to critical ground support events for cabin door closing flights are forcibly retained. The directed causal edges from critical ground support events for cabin door closing flights to critical ground support events for passenger boarding bridge evacuation flights are forcibly retained. The directed causal edges from critical ground support events for passenger boarding bridge evacuation flights to critical ground support events for aircraft pushback and wheel chock removal flights are forcibly retained.

[0056] Subsequently, the causal structure feasible region matrix is ​​generated. Based on the aforementioned constraint set, a binary feasible region matrix of dimension m×m is generated, where m is the total number of critical events and influencing factors for flight ground support. The formula for calculating the feasible region matrix is: ; In the formula, Let be the feasible region matrix of the causal structure; i is the starting variable index of the causal edge, ranging from 1 to m; j is the ending variable index of the causal edge, ranging from 1 to m. A matrix element of 1 indicates that the corresponding directed causal edge is allowed, and a matrix element of 0 indicates that the corresponding directed causal edge is prohibited, thus defining the search boundary of the causal structure.

[0057] Finally, a loop closure check is performed on the generated feasible region matrix to ensure that the directed graph corresponding to the matrix does not contain any cyclic causal paths, meeting the basic requirements of a directed acyclic graph. The civil aviation business prior constraint set and causal structure feasible region matrix generated in this step will be directly input into the causal discovery algorithm in step three as hard constraints in the algorithm's search process.

[0058] Step 3, Constrained NOTEARS-MCP Algorithm for Causal Structure Learning: This step employs the NOTEARS-MCP algorithm (NOTEARS, Non-combinatorial Optimization via TraceExponential and Augmented lagRangian for Structure learning; MCP, Minimax Concave Penalty) with adjacency matrix smoothing constraints to complete causal structure learning. This algorithm transforms discrete directed acyclic graph learning into a continuous optimization problem, significantly reducing computational complexity in high-dimensional, small-sample civil aviation scenarios. Simultaneously, it effectively suppresses the generation of spurious causal edges through non-convex penalties and neighborhood constraints, addressing the core pain points of conventional algorithms in civil aviation scenarios, such as temporal errors and spurious causal relationships. The causal structure learning and causal effect quantification process is as follows: Figure 3 As shown.

[0059] The core objective of the algorithm is to learn the m×m dimensional adjacency matrix. Matrix elements represents the weight of the causal edge from variable i to variable j; a non-zero weight indicates a causal relationship between the two variables. The algorithm transforms the learning of a directed acyclic graph into a constrained continuous optimization problem, avoiding the high complexity of traditional discrete search methods.

[0060] The objective function of the algorithm is then constructed. It consists of four parts: a data fitting term, a non-convex penalty term, a priori constraint term, and an acyclic constraint term. The complete formula is as follows: In the formula, n is the number of samples in the standardized quantization reference dataset; X is the n×m dimensional sample matrix formed by the standardized quantization reference dataset. Let Frobenius norm be the matrix; is the regularization coefficient for the MCP penalty term, with a fixed value of 0.02; This is the non-convex penalty function of the MCP, used to compress the weights of spurious causal edges to zero; The concavity parameter for MCP penalty is fixed at 3; The weighting coefficient for the prior constraint term is fixed at 100. This is the Hadamard product operation for matrices, which is the multiplication of corresponding elements of two matrices. This is the penalty coefficient for the acyclic constraint term, initially set to 1, and increasing exponentially during iteration. is a smoothing constraint function for directed acyclic graphs, used to ensure that the learned adjacency matrix corresponds to an acyclic graph.

[0061] The formula for calculating the smoothness constraint function of a directed acyclic graph is as follows: In the formula, This is the trace operation of a matrix, which is the sum of the elements on the main diagonal of the matrix; This refers to the exponentiation operation of a matrix.

[0062] The next step is to complete the training process of the algorithm. The first step is to initialize the adjacency matrix. The matrix is ​​all zeros; the penalty coefficients are initialized. Set the maximum number of iterations to 1000 and the convergence threshold to [value missing]. The second step uses the ADMM (Alternating Direction Method of Multipliers) optimization algorithm to solve for the minimum objective function under the current penalty coefficient, thus obtaining the updated adjacency matrix. The third step is to calculate the value of the acyclic constraint function. ,like If the iteration stops and the current adjacency matrix is ​​not satisfied, then the iteration stops and the current adjacency matrix is ​​output. If the convergence condition is not met, then the iteration stops. Multiply by 10 and return to step two to continue iterating; in step four, if the iteration count reaches the maximum number of iterations and still has not converged, output the adjacency matrix with the smallest acyclic constraint function value during the iteration process.

[0063] The algorithm is then applied to the adjacency matrix obtained from the training. Thresholding is performed, setting elements with absolute values ​​less than 0.1 to zero to obtain the final sparse causal adjacency matrix. Based on this matrix, a causal graph of flight ground support events is drawn to clarify the causal interaction paths between all variables.

[0064] This step incorporates a bidirectional interaction mechanism with the subsequent step five. The causal adjacency matrix output in this step... This will serve as the basis for selecting covariates in the entropy balance weighted causal effect estimation algorithm in step five. Simultaneously, the average intervention effect output from step five will inversely correct the penalty term coefficient of this algorithm, achieving synergistic optimization of causal structure and causal effect. The interaction formula is as follows: In the formula, The updated regularization coefficients; is the initial regularization coefficient; k is the total number of effective intervention intervals; This represents the average intervention effect corresponding to the i-th intervention interval output in step five.

[0065] Step Four: Heterogeneous Interval Division of Influencing Factors and Generation of Effective Intervention Sets: The core function of this step is to divide the influencing factors of civil aviation ground support into heterogeneous intervention intervals, addressing the problem that conventional binary interventions cannot characterize the differentiated causal effects across different time intervals. Civil aviation ground support control actions involve continuous time adjustments; the impact of ground staff arriving 5 minutes or 15 minutes early on support duration differs significantly. Therefore, interval division is essential for achieving refined quantification of causal effects.

[0066] First, the initial division of the single-factor intervention interval was completed. For the standardized quantitative reference quantity of each flight ground support influencing factor, the equal frequency quartile division method was used to complete the initial interval division. The continuous time reference quantity was divided into 4 initial intervals with balanced sample sizes to avoid the interference of extreme values ​​on the division results and ensure that each interval has a sufficient sample size to support the subsequent causal effect estimation.

[0067] Subsequently, the business adaptation adjustment of the interval boundaries was completed. In combination with the operation time scale of flight ground support, the boundaries of the initial intervals were adjusted so that the boundaries of all intervals are integer multiples of 0.5 minutes, which matches the minimum control time unit of civil aviation support operations. This ensures that the divided intervals conform to the control granularity of the actual airport scheduling and avoids interval division results that are theoretically reasonable but practically unenforceable.

[0068] Next, a dual compliance check was performed on the intervention intervals. Two checks were conducted on all adjusted single-factor intervention intervals, and intervals that did not meet the requirements were removed. The first check was the sample size check. If the sample size of an intervention interval accounted for less than 5% of the total sample size, the interval was directly removed to avoid causal effect estimation bias caused by small samples. The second check was the covariate overlap check. The standardized mean difference method was used to calculate the standardized mean difference between the interval and other intervals on key covariates such as flight type, time period, and gate type. If the standardized mean difference was greater than or equal to 0.2, the overlap was deemed insufficient, and the interval was removed to ensure that the positive assumption of causal effect estimation was valid. This is the core prerequisite for causal inference.

[0069] Finally, the construction of multi-factor joint intervention intervals was completed. For multiple flight ground support influencing factors corresponding to key ground support events of the same flight, the verified single-factor intervention intervals were combined by Cartesian product to form the initial multi-factor joint intervention intervals. The sample size of the initial joint intervals was verified, and combinations in which the sample size of the joint interval accounted for less than 3% of the total sample size were removed. Finally, the effective intervention interval set corresponding to each flight ground support key event was formed.

[0070] This step incorporates a two-way interaction mechanism with the subsequent step eight. The set of effective intervention intervals output in this step will serve as the action space input for the path integral control algorithm in step eight. Simultaneously, the optimal policy action access frequency output in step eight will inversely correct the interval partitioning granularity in this step, refining the partitioning of frequently selected high-value intervals to further enhance the optimization space of the policy. The interaction formula is as follows: In the formula, The granularity of the updated interval division; Define the initial interval granularity; The access frequency for each action output in step eight; For variance calculation; This is for averaging.

[0071] Step 5, Entropy Balance Weighted Causal Effect Quantification and Intensity Ranking: This step uses an entropy balance weighted causal effect estimation algorithm to quantify the average intervention effect. This algorithm does not require fitting a propensity score model and can directly achieve a perfect match between the covariate distributions of the intervention group and the control group through entropy balance constraints. In the civil aviation scenario with small samples and high-dimensional covariates, its robustness is much higher than that of conventional methods, effectively avoiding estimation bias caused by model misspecification.

[0072] The core objective of the algorithm is to calculate the average intervention effect corresponding to each effective intervention interval, quantifying the causal impact of the intervention action on the execution time of critical ground support events for the target flight, and providing a quantitative causal basis for subsequent strategy optimization. The algorithm uses entropy balance weighting to ensure a perfect match between the covariate moments of the intervention group and the control group, eliminating confounding bias and obtaining unbiased causal effect estimation results.

[0073] Subsequently, the entropy balance weight optimization model was constructed. For each effective intervention interval, samples falling within that interval were designated as the intervention group, and the remaining samples as the control group. The optimization objective of the algorithm is to minimize the entropy loss of the control group sample weights while satisfying the covariate moment balance constraint. The complete formula is: ; The constraints are: ; ; ; In the formula, The entropy balance weight is the i-th sample in the control group; This is the index set for the control group samples; This is the set of indices for the intervention group samples; Let k be the value of the covariate for the i-th sample. The covariate is determined by the causal adjacency matrix output in step three, and only variables that have a causal relationship with the target event are retained. is the sample size of the intervention group; k is the index of the covariate, ranging from 1 to the total number of covariates.

[0074] The next step is to complete the training process of the algorithm. The first step is to determine the intervention group and control group samples for each effective intervention interval and extract the corresponding covariate matrix. The second step is to use the Newton-Raphson algorithm to solve the above-mentioned constrained optimization problem and obtain the optimal entropy balance weights for the control group samples. The third step, based on the optimal weights, calculates the mean difference between the standardized quantitative reference values ​​of the target events in the intervention group and the weighted control group, thus obtaining the average intervention effect corresponding to the intervention interval. The calculation formula is as follows: In the formula, This represents the average intervention effect corresponding to the intervention interval, in minutes; This serves as a standardized quantitative reference value for the critical ground support event for the target flight corresponding to the i-th sample. A positive average intervention effect indicates that the intervention interval will prolong the execution time of the critical ground support event for the target flight; the larger the value, the stronger the impact. A negative average intervention effect indicates that the intervention interval will shorten the execution time of the critical ground support event for the target flight; the larger the absolute value, the stronger the impact. The fourth step repeats the above process for all effective intervention intervals to obtain the average intervention effect for all intervals.

[0075] The algorithm was then applied. For each critical ground support event for a flight, all effective intervention intervals were sorted from largest to smallest based on the absolute value of the average intervention effect, forming a causal strength ranking table. This provided a priority basis for subsequent action space compression and initial strategy construction.

[0076] This step incorporates a bidirectional interaction mechanism with the preceding and following steps. On one hand, it interacts in reverse with the NOTEARS-MCP algorithm in step three; the average intervention effect output in this step will correct the regularization coefficient of the algorithm in step three, achieving synergistic optimization of causal structure and causal effect. On the other hand, it interacts bidirectionally with the path integral control algorithm in step eight; the average intervention effect output in this step will serve as the weight coefficient of the reward function in the algorithm in step eight, while the optimal policy action access frequency output in step eight will correct the covariate selection weight of this algorithm. The interaction formula is as follows: In the formula, The weights of the updated k-th covariate; These are the initial covariate weights; This is the access frequency of the action related to the k-th covariate, as output in step eight.

[0077] Step Six: Markov Decision Process Modeling and Space Compression for the Safeguard Process: The core function of this step is to transform the discrete, multi-stage flight ground support process into a mathematically solvable sequential decision model. It serves as a crucial bridge between causal analysis and strategy optimization, enabling a leap from understanding which factors affect efficiency to determining how to manage them to improve on-time performance. The causal reinforcement-based global optimal strategy solution process is as follows: Figure 4 As shown.

[0078] First, the Markov property of flight ground handling was verified. The verification results show that the entire flight ground handling process conforms to the Markov property, that is, given the execution state of the current critical event in flight ground handling and the states of all preceding events, the conditional probability distribution of the future event state depends only on the state of the current event. The core logic of this property is that the execution state of each handling link is only related to the completion state of the previous link, and has no direct relationship with earlier links. For example, the efficiency of cabin cleaning is only related to the completion state of passenger disembarkation, and has no direct relationship with previous wheel chock placement and jet bridge docking. This fully conforms to the Markov property of no aftereffect. After the verification was passed, the Markov chain model of the entire process was completed.

[0079] Subsequently, the scenario mapping of the six core elements of the Markov decision-making process for flight ground support was completed, and the complete definitions are as follows: The first element is the intelligent agent, defined as a software program that executes time-constrained decisions and is responsible for the selection of actions and optimization of strategies throughout the entire process; the second element is the environment, defined as the business logic, fixed temporal sequence, and real-time execution status of events throughout the flight ground support process, responsible for providing feedback on state transitions and rewards after action execution; the third element is the state, defined as the critical event in flight ground support at the current moment, and the standardized quantitative reference value of the event, i.e., the execution duration of the event; the fourth element is the action, defined as the combination of effective intervention intervals that can be selected under the current critical event in flight ground support, i.e., the set of effective intervention intervals output in step four; the fifth element is the strategy, defined as the mapping relationship from state to action, i.e., the conditional probability of selecting the corresponding action in the current state; the sixth element is the reward, defined as the degree of consistency between the planned departure time and the actual departure time of the flight. The higher the degree of consistency, the greater the reward value. The complete formula of the reward function is: In the formula, This is an instant reward value, ranging from 0 to 10; ; This refers to the actual departure time of the flight; The times listed are the scheduled departure times for flights, all in minutes. ; .

[0080] Finally, state-action space compression based on causal strength was performed. For the action space of critical ground support events for each flight, based on the causal strength ranking table, low-impact actions with an average intervention effect of less than 0.1 minutes were eliminated, retaining only effective actions with high causal strength. This compressed the action search space and reduced the computational complexity of subsequent optimizations. The core significance of this operation is that if the action space is too large, the subsequent reinforcement learning algorithm will converge very slowly, or even fail to converge. By eliminating invalid actions through causal priors, the convergence efficiency and policy quality of the algorithm can be significantly improved.

[0081] Step 7, Construction of the Initial Behavior Policy Set under Causal Constraints: The core function of this step is to provide a high-quality initial policy for the subsequent path integral control algorithm, avoiding the problems of slow convergence and getting trapped in local optima caused by learning from random policies. The initial policy is not randomly generated, but constructed based on the preceding causal effect results, which is equivalent to injecting causal priors into the algorithm, allowing it to focus on effective control actions from the beginning, significantly improving the efficiency and quality of policy optimization.

[0082] First, the initial causal weights are calculated. For all effective actions in each state, the initial causal weights are calculated based on the average intervention effect corresponding to that action. The formula is as follows: In the formula, Let be the initial causal weights for action a in state s; The average intervention effect corresponding to action a; To prevent smoothing terms with a denominator of 0, the core logic of this formula is that the larger the absolute value of the average intervention effect, the stronger the impact of the action on the duration of protection. If the average intervention effect is positive, it means that the action will extend the duration of protection, and the lower the initial weight, the lower the probability of being selected. If the average intervention effect is negative, it means that the action will shorten the duration of protection, and the higher the initial weight, the higher the probability of being selected. This fully conforms to the core logic of causal control.

[0083] Subsequently, state reward correction is performed, introducing a correction function based on action execution rewards to adjust the initial causal weights. The corrected weight formula is as follows: In the formula, The corrected weights; The correction factor is fixed. The benefit of state change after performing action a in state s is calculated if the execution time of the target flight's ground support critical event is shortened compared to the baseline value after performing action a. ,and ,but If the execution time is neither shortened nor extended, then The core purpose of this revision is to balance theoretical causal effects with practical feasibility, avoiding the problem of assigning excessive weight to actions that are theoretically effective but do not comply with airport scheduling rules or are impractical in actual implementation.

[0084] Next, the initial policy probabilities are normalized. The corrected weights are normalized to obtain the selection probability of each action in each state, forming the initial behavioral policy set, as shown in the formula: In the formula, Let be the initial policy probability for choosing action a in state s; Let s be the set of all valid actions in state s. for Any of the alternative actions in the selection.

[0085] Finally, the feasibility of the initial strategy is verified. The initial behavioral strategy set is verified to ensure that the sum of the selection probabilities of all actions in each state is 1, and all probability values ​​are within the range of 0 to 1. After the verification is passed, the final initial behavioral strategy set is determined.

[0086] Step 8: Solving for the Global Optimal Policy using the Path Integral Control Algorithm: This step employs a path integral control algorithm based on a linearly solvable Markov decision process to solve for the global optimal policy. Compared to conventional reinforcement learning algorithms such as Q-learning and DQN, this algorithm transforms the nonlinear Markov decision process into a linearly solvable problem. It does not require a large number of sample iterations, directly calculates the optimal policy through path integrals, has a fast convergence speed, and can seamlessly incorporate preceding causal priors, avoiding getting trapped in local optima. This makes it very suitable for scenarios with high safety requirements and limited sample sizes, such as civil aviation.

[0087] The core objective of the algorithm is to find the optimal stochastic strategy for the entire flight ground support process, maximize the expected cumulative reward from the initial state to the final state, and finally obtain the globally optimal time-constrained strategy for the entire flight ground support process.

[0088] Subsequently, the transformation into a linearly solvable Markov decision process is completed. Based on the constructed Markov decision process, the state transition probability is combined with the action cost to construct a linearly solvable dynamic programming equation. First, the immediate cost of a state is defined as the negative immediate reward, i.e. The core dynamic programming equation of the algorithm is: In the formula, Let s be the value function of state s; This is a discount factor used to balance immediate costs with future long-term costs; Let be the probability of transitioning to state s' after performing action a in state s. This probability is constrained by the causal adjacency matrix output in step three, allowing only state transitions that conform to causal paths.

[0089] Introducing exponential transformation ,in As the inverse temperature coefficient, fixed at 0.5, the above nonlinear dynamic programming equation is transformed into a linear equation, as shown in the formula: In the formula, The initial behavior policy probabilities generated in step seven. The core function of this exponential transformation is to convert the minimization operation of the Bellman equation into a linear summation operation, thus realizing the linear solvability of the Markov decision process and avoiding the high computational cost of the iterative Bellman update in conventional algorithms.

[0090] The next step is to complete the training process of the algorithm. The first step is to initialize the state value function. As a vector of all zeros, initialize Given a vector of all 1s, a maximum number of iterations is set to 10000, and a convergence threshold is... The second step is based on the current situation. Calculate the optimal action probability for each state using the following formula: The third step is based on the updated strategy. Recalculate the state value function and update Step 4 Calculation If the maximum change is less than the convergence threshold, the iteration stops and the current optimal policy is output. If the convergence condition is not met, the iteration returns to the second step to continue. In the fifth step, if the iteration count reaches the maximum number of iterations and convergence is still not achieved, the policy with the largest cumulative reward during the iteration process is output.

[0091] The algorithm is then applied. Based on the converged optimal strategy, the action with the highest probability is selected for each state, forming a complete sequence of actions from the initial state to the final state, which is the globally optimal time constraint strategy.

[0092] This step designs a bidirectional interactive closed loop with the previous steps. The causal adjacency matrix output in step three constrains the state transition probability matrix of this algorithm. Simultaneously, the state transition frequency of the optimal policy output by this algorithm inversely corrects the adjacency matrix weights in step three. The interaction formula is as follows: In the formula, The transition frequency from state i to state j is used; the average intervention effect output in step five is used as the weight coefficient of the immediate cost function of this algorithm. At the same time, the optimal strategy action access frequency output by this algorithm corrects the covariate selection weight in step five; the optimal strategy action access frequency output by this algorithm corrects the interval partitioning granularity in step four, forming a complete closed loop from causal analysis to strategy optimization, and then from strategy optimization to reverse optimization of causal analysis.

[0093] Step 9, Optimal Strategy Robustness Perturbation Verification: The core function of this step is to verify the stability of the strategy in actual operation, ensuring that the strategy is not only effective under ideal historical data, but also remains stable under uncertain scenarios in actual civil aviation operations, avoiding the problem of strategy overfitting to historical data and poor actual implementation results.

[0094] First, the disturbance verification scheme was designed. For the average intervention effect value corresponding to all actions in the globally optimal time constraint strategy, disturbances of +10% and -10% were applied respectively to obtain the average intervention effect value after disturbance. A disturbance amplitude of ±10% was chosen, which is the normal fluctuation range of time deviation in actual civil aviation operations, effectively simulating the uncertainties in actual operation.

[0095] Subsequently, the stability of the strategy is verified. Based on the average intervention effect value after the perturbation, the causal strength ranking of each action is recalculated, and the ranking changes before and after the perturbation are compared. If the ranking change rate is less than 10%, the strategy is deemed to have passed the robustness verification and is the final effective strategy; if the ranking change rate is greater than or equal to 10%, the strategy is deemed to be insufficiently robust, triggering the feedback optimization process.

[0096] Next, the feedback optimization process is completed. For critical events in flight ground support with insufficient robustness, the verification results are sent back to step four, and the intervention interval division is re-executed. The interval division granularity is adjusted from the quartile to the quintile to refine the interval division accuracy. Then, the entire process from step four to step nine is repeated until the output globally optimal time constraint strategy passes the robustness verification.

[0097] Step 10, Output and Implementation Mapping of Flight Ground Support Event Time Control Requirements: The core function of this step is to transform the mathematically optimal strategy into control requirements that frontline airport support personnel can directly execute, thus achieving a closed loop from theoretical model to actual implementation and completing the final output of the entire method.

[0098] First, the mapping from strategy to control requirements is completed. Each action in the final global optimal time constraint strategy is transformed into the landing time control requirements of the corresponding flight ground support influencing factors. The control requirements clearly define the time difference control range of the corresponding influencing factors, which corresponds one-to-one with the calculation formula of the standardized quantitative reference quantity defined in step one, to ensure the executability and traceability of the control requirements.

[0099] Subsequently, the integration of the full-process control requirements was completed. According to the time sequence of the entire flight ground support process, the control requirements of all key events of flight ground support were integrated to form a complete and time-coherent flight ground support full-process time control manual, which clarifies the pre-control nodes, time control requirements and responsible entities of each link, so that the airport's operation command department and ground support department can use it directly.

[0100] The final output, a manual outlining the time control requirements for the entire flight ground support process, represents the final implementation of the entire methodology. It can be directly applied to the operation and management of flight ground support at airports, thereby improving the on-time departure rate of flights.

[0101] In some embodiments, for peak operation scenarios of hub airports with an annual passenger throughput of over ten million in China, the typical parameter values ​​for each core algorithm and component of this solution are as follows: 1. Parameters related to the NOTEARS-MCP algorithm with adjacency matrix smoothing constraints: MCP penalty term regularization coefficient The value is fixed at 0.02, which is the MCP penalty concavity parameter. The value is fixed at 3, and the weight coefficient of the prior constraint term is... The initial penalty coefficient for the acyclic constraint term is fixed at 100. The initial value is 1, and it is multiplied by 10 in each iteration. The maximum number of iterations is set to 1000, and the convergence threshold is set to... The threshold for adjacency matrix thresholding is set to 0.1; 2. Parameters of the causal effect estimation algorithm based on entropy balance weighting: The covariate balance order is set to 1, i.e., matching the covariate means of the intervention group and the control group; the maximum number of iterations of the Newton-Raphson algorithm is set to 500; and the convergence threshold is set to... ; 3. Parameters related to the path integral control algorithm based on a linearly solvable Markov decision process: discount factor The value is fixed at 0.95, representing the inverse temperature coefficient. The value is fixed at 0.5, the maximum number of iterations is set to 10000, and the convergence threshold is set to... ; 4. Interval partitioning and validation parameters: The initial interval partitioning adopts the quartile partitioning of equal frequency. The sample size threshold for single-factor intervention intervals is set to 5% of the total sample size. The sample size threshold for multi-factor joint intervention intervals is set to 3% of the total sample size. The standardized mean difference threshold for covariate overlap validation is set to 0.2. 5. Robustness perturbation verification parameters: The perturbation amplitude of the average intervention effect is set to ±10%, and the ranking change rate threshold for strategy stability verification is set to 10%.

[0102] The parameter values ​​in this embodiment are adapted to peak scenarios at typical hub airports. The parameters can be adjusted regularly according to different airport sizes, flight densities, and operating scenarios.

[0103] In some embodiments, taking a domestic Boeing 737-800 flight at a domestic hub airport as an example, the following is an example of the full-process calculation and application of this solution for a single flight: Step 1: Obtain the raw data of the entire historical support process for this flight. Based on the hierarchical analysis system of this plan, identify the nine key ground support events and corresponding influencing factors for this flight. Calculate all standardized quantitative reference quantities based on the time node data. For example, if the arrival time of the wheel chock operation ground crew at the gate is 3 minutes earlier than the flight's arrival time, what is the standardized quantitative reference quantity for the arrival of the wheel chock operation ground crew at the gate? The calculated value is 3 minutes; the aircraft gate check completion time for this flight is 2 minutes earlier than the flight's arrival time at the gate, corresponding to the standardized quantitative reference value of the aircraft's arrival status at the gate. The calculation time is 2 minutes, and so on, to complete the calculation of all reference quantities. After being truncated by 3 times the standard difference constant, they are included in the standardized quantitative reference quantity dataset. Step 2: Based on the civil aviation ground support business specifications and the support process sequence of the flight, construct the corresponding civil aviation business prior constraint set, generate the causal structure feasible domain matrix of the scenario to which the flight belongs, and clarify the directions of prohibited and mandatory causal edges; Step 3: Using the standardized quantitative reference dataset of the scenario to which the flight belongs as input, and the prior constraint set and feasible region matrix as constraints, the NOTEARS-MCP algorithm with adjacency matrix smoothing constraint is used to solve the problem and obtain the sparse causal adjacency matrix in the scenario. The core causal relationships are identified, such as the arrival status of ground staff for wheel chock operation having a significant causal impact on the key event of aircraft wheel chock placement, and the arrival status of cabin cleaning staff having a significant causal impact on the key event of cabin cleaning. The corresponding causal graph is then drawn. Step 4: Based on the standardized quantitative reference dataset and the sparse causal adjacency matrix, the reference values ​​of each influencing factor are divided into equal-frequency quartile intervals. For example, the reference value of the arrival of ground staff at the aircraft position for wheel chock operation is divided into four initial intervals: (0,2], (2,4], (4,6], and (6,8]. After sample size verification and covariate overlap verification, the effective intervention interval set of the influencing factor is determined. Step 5: Using an entropy balance weighted causal effect estimation algorithm, calculate the average intervention effect corresponding to each effective intervention interval. For example, the average intervention effect calculated for the (0,2] interval of the situation where ground crew arrives at the aircraft stand for wheel chock operation is 2.9 minutes. That is, the intervention action in this interval will increase the execution time of the critical event of placing the aircraft wheel chock by an average of 2.9 minutes. The average intervention effect calculation for all intervals is completed in this way, and a causal strength ranking table is generated. Step 6: Based on the sparse causal adjacency matrix, average intervention effect, and causal intensity ranking table, construct a Markov decision process model for flight ground support in this scenario, eliminate low-impact actions with an average intervention effect absolute value of less than 0.1 minutes, and complete the state-action space compression. Step 7: Based on the average intervention effect and causal strength ranking table, calculate the initial causal weight of each action, and after state change benefit correction and normalization, generate the initial behavioral strategy set for this scenario. Step 8: Using the path integral control algorithm based on a linearly solvable Markov decision process, the global optimal time constraint strategy for the scenario to which the flight belongs is obtained. For example, for the critical event of aircraft wheel chock placement, the intervention interval selected by the optimal strategy is (2,4], which means that the ground staff operating the wheel chocks must arrive at the gate 2 to 4 minutes earlier than the flight arrival time. Step 9: Perform an average intervention effect perturbation verification of the obtained globally optimal time constraint strategy with an amplitude of ±10%. The verification results show that the ranking change rate is 3.2%, which is less than the threshold of 10%. The strategy passes the robustness verification and is determined to be the final globally optimal time constraint strategy. Step 10: Transform the final global optimal time constraint strategy into the landing time control requirements corresponding to the flight, clarify the pre-control nodes, time control intervals and responsible positions for each support link, and output them to the airport operations command department and ground support department for execution.

[0104] In some embodiments, to address the industry pain points of insufficient sample size leading to causal structure learning bias and poor prior constraint adaptability in civil aviation emergency operation scenarios (such as extreme weather, flight diversion, and temporary adjustment of gate positions), a meta-learning-driven dynamic prior constraint adaptive update and small sample causal structure adaptation step is added between step two and step three. At the same time, the NOTEARS-MCP algorithm is optimized and reconstructed by meta-constraints to form a dual-loop causal structure learning framework.

[0105] The specific implementation method is as follows: 1. Construction of Prior Knowledge Meta-Feature Database: Based on historical data of different civil aviation operation scenarios (normal peak, extreme weather, flight diversion, night operation, and wide-body aircraft support), prior constraint sets, causal structure feasible domain matrices, and core causal edge weight distributions are extracted for each scenario to construct a meta-knowledge database containing scenario meta-features, prior constraint meta-parameters, and causal structure meta-weights. The meta-feature vector of each scenario contains five core dimensions: flight density, gate turnover rate, weather impact level, flight diversion rate, and wide-body aircraft proportion. The matching degree between the target scenario and each scenario in the meta-feature database is calculated using cosine similarity.

[0106] 2. Meta-learning adaptation of dynamic prior constraints: For a new target operating scenario, if the effective sample size of the scenario is lower than a preset threshold (e.g., less than 500 samples per scenario), then based on the scenario matching degree, the prior constraint meta-parameters of the top 3 similar scenarios are extracted from the meta-knowledge database. Meta-weighted fusion is then used to generate a dynamic prior constraint set and an initial feasible region matrix adapted to the target scenario. The meta-weighted fusion formula is: In the formula, This is the dynamically generated feasible region matrix for the target scenario. Let the matching degree weight of the k-th similar scene satisfy... , This is the feasible domain matrix for the kth similar scenario; at the same time, for the unique business rules of the target scenario, the fused feasible domain matrix is ​​supplemented with mandatory constraints and verified with prohibitive constraints to ensure compliance with the time sequence logic of civil aviation business.

[0107] 3. Meta-constraint Enhanced NOTEARS-MCP Algorithm Reconstruction: A meta-regularization constraint term is added to the original algorithm's objective function. The causal structure meta-weights of similar scenarios are used as prior information to constrain the adjacency matrix learning process of the target scenario. The reconstructed objective function is: ; In the formula, is the regularization coefficient, with a fixed value of 0.05. This is the meta-weight matrix of the causal structure after fusing similar scenarios. The reconstructed algorithm can significantly reduce the probability of generating false causal edges in small sample scenarios by leveraging meta-prior information, thereby improving the accuracy and convergence speed of causal structure learning.

[0108] 4. Dynamic prior online update closed loop: When the cumulative number of new samples in the target scene exceeds the preset threshold, causal structure learning is re-executed based on the new samples to update the meta-features and prior constraint meta-parameters of the scene, and the meta-knowledge database is updated synchronously to achieve continuous adaptive optimization of prior constraints.

[0109] This embodiment breaks through the inherent ideas of fixed prior constraints and reliance on large sample causal learning in the existing technology. It deeply integrates meta-learning and causal structure learning, solves the industry pain point of small sample causal modeling in civil aviation emergency scenarios, and further improves the adaptability and accuracy of the strategy in complex dynamic scenarios.

[0110] In some embodiments, to address the problem that only the direct causal effect of intervention actions is quantified while ignoring the transmission influence of chain mediation effects in multiple links of the flight ground support serial process, resulting in the strategy optimization failing to cover the indirect causal gain of the entire process and insufficient global optimization capability, the entropy balance weighted causal effect estimation algorithm in step five is extended by adding a hierarchical quantification of chain mediation causal effects and a heterogeneity moderating effect modeling step, so as to achieve unbiased quantification of direct effects, mediation effects and total effects in all dimensions.

[0111] The specific implementation method is as follows: 1. Identification of Causal Chain Mediation Paths: Based on the sparse causal adjacency matrix and causal graph output in step three, for each target key event, identify chain mediation paths with the event as the outcome, intervention factors as exposures, and intermediate safeguard key events as mediating variables; for example, for the cabin door closing key event, identify a three-level chain mediation path: cabin cleaning personnel arrival status → cabin cleaning time → departing passenger boarding efficiency → cabin door closing time, and clarify the mediating variables and causal dependencies at each level of the path.

[0112] 2. Hierarchical Entropy Balance Weighted Decomposition of Mediation Effects: For the identified chain-like mediation paths, a hierarchical entropy balance weighted model is constructed. Covariate balancing is performed for each level of mediation to eliminate confounding biases and achieve unbiased decomposition of direct and mediation effects. Specifically: The first step is to perform entropy balance weighting on the exposure variable (intervention factor) and the outcome variable (target key event) to calculate the total causal effect. ; The second step is to address the mediator variables at each level. (k is the mediator number), using the exposure variable as the intervention and the mediator variable as the outcome, entropy balance weighting is performed to calculate the causal effect of the exposure on the mediator variable. ; The third step involves performing a joint entropy balance weighting, with the exposure variable and all mediating variables as the joint intervention and the outcome variable as the target. Under the condition of controlling for mediating variables, the direct causal effect of the exposure on the outcome is calculated. And the causal effect of each level of mediating variable on the outcome. ; The fourth step is to calculate the specific mediation effect of each level of the path based on the product rule of chain mediation effects. and the overall mediating effect Where K is the total number of intermediaries, satisfying .

[0113] 3. Modeling of Heterogeneous Moderating Effects: Introducing aircraft type, gate type, operating time, and flight density as moderating variables, a hierarchical model of moderating effects is constructed to quantify the heterogeneity differences between direct and mediating effects at different levels of moderating variables. For each level of moderating variable, the above-mentioned mediating effect decomposition is performed to generate a quantitative table of mediating effects in different scenarios, clarifying the core mediating transmission path in different scenarios.

[0114] 4. Process Integration: The decomposed total causal effect, mediating effect, and direct effect are synchronously updated to the causal intensity ranking table of the intervention interval. In the state-action space compression step of step six, in addition to considering the total causal effect, actions with an absolute value of mediating effect greater than a preset threshold (0.2 minutes) are additionally retained to avoid eliminating intervention actions with significant indirect gains. In the path integral control algorithm of step eight, the mediating effect is included in the constraint condition of the state transition probability to ensure that the state transition conforms to the transmission logic of the chain-mediated causal path.

[0115] This embodiment breaks through the inherent limitation of existing civil aviation support optimization technologies that only focus on direct causal relationships between variables. Targeting the characteristics of the serial process of flight support, it constructs a hierarchical entropy balance weighted chain mediation effect quantification method, which accurately characterizes the indirect transmission gain of intervention actions in the entire process. This fundamentally avoids the problem of core intervention actions being mistakenly eliminated and global optimization strategies getting stuck in local optima due to ignoring mediation effects, and further enhances the global collaborative optimization capability.

[0116] In some embodiments, to address the industry pain point that the single-objective reward function only focuses on the on-time departure rate of flights and does not incorporate core constraints such as civil aviation safety hard constraints, the balance of ground support personnel workload, and equipment scheduling costs, resulting in safety risks and poor implementability of the optimization strategy, the Markov decision process model in step six and the path integral control algorithm in step eight are reconstructed. A hierarchical reward mechanism with multi-objective opportunity constraints of safety, efficiency, and workload is added to achieve multi-objective collaborative optimization of on-time performance improvement, safety compliance, and workload balance.

[0117] The specific implementation method is as follows: 1. Opportunity Constraint Modeling of Hard Constraints for Civil Aviation Safety: An opportunity constraint programming model is constructed to address the statutory safety requirements of each stage of flight ground support, ensuring that the optimization strategy does not exceed the minimum duration constraint for safe operations while meeting pre-set confidence levels. Specifically, for each critical support event... Set a legally mandated minimum safe operating time. Construct opportunity constraints: In the formula, To constrain the maximum permissible probability of violation, a fixed value of 0.01 is set, ensuring that the minimum safe duration requirement is met with a probability of over 99%. This opportunity constraint will serve as a hard constraint on action selection during the Markov decision-making process, prohibiting the selection of intervention actions that would cause the probability of safety constraint violation to exceed the limit.

[0118] 2. Construction of a multi-objective hierarchical reward function: The single-objective reward function described above is reconstructed into a three-layer progressive reward function, namely, a security compliance layer, an efficiency improvement layer, and a load balancing layer. The rewards of each layer are weighted and merged to form the total reward function, as follows: Security compliance layer rewards If the action performed in the current state satisfies the above safety opportunity constraints, If the restrictions are violated, As a penalty, the action is directly rejected. Efficiency Improvement Layer Rewards Based on the above calculation of departure time consistency, the formula remains the same as above, with a value range of 0~10; Load balancing layer rewards Based on the calculation of the workload balance of ground support personnel, the workload is measured by the number of flights supported by personnel per unit time and the standard deviation of the operation time. The formula is as follows: ; In the formula, This represents the standard deviation of the working time for each support team under the current operational conditions. The maximum permissible standard deviation is preset. The value ranges from 0 to 5; the more balanced the load, the higher the reward value. The total reward function is: Safety is the top priority, while also taking into account efficiency and load balancing.

[0119] 3. Optimization of Constraint-Enhanced Path Integral Control Algorithm: In the aforementioned path integral control algorithm, safety opportunity constraints are used as pre-filtering conditions for action selection. In each iterative step of solving for the optimal action probability, actions violating safety constraints are first eliminated. Simultaneously, the multi-objective total reward function is used as the basis for calculating the algorithm's immediate cost function, i.e. This ensures that the optimal strategy for solving the algorithm simultaneously satisfies the three core objectives of safety, efficiency, and load.

[0120] 4. Integration with the robustness verification above: In the robustness disturbance verification in step nine above, a new safety constraint violation rate verification item is added. For the perturbed strategy, the probability of violation of safety constraints is calculated. If the violation probability exceeds 1%, the strategy verification is deemed to have failed and is fed back to step four to adjust the division of the intervention interval, eliminating the interval that will break the safety constraints, so as to ensure that the final strategy always meets the civil aviation safety compliance requirements.

[0121] This embodiment breaks through the inherent technical bias of existing civil aviation support optimization technologies that prioritize efficiency over safety and ignore feasibility. It incorporates statutory safety constraints of civil aviation into the reinforcement learning framework through opportunity constraint modeling, and constructs a multi-objective hierarchical reward mechanism that prioritizes safety. This not only ensures the safety and compliance of the optimization strategy, but also takes into account the workload balance of front-line support personnel, and greatly improves the feasibility of the strategy.

[0122] In some embodiments, to address the issues of coordination conflicts, mismatched scheduling responsibilities, and significant obstacles to strategy implementation in the distributed execution scenario of multiple support entities (ground support, jet bridges, cleaning, aircraft maintenance, and cargo) at airports, the above-mentioned centralized decision optimization is supplemented by a causal potential game modeling step for multi-entity distributed collaborative intervention between steps four and seven. This extends the above-mentioned centralized Markov decision process into a multi-agent causal potential game model, ensuring that the distributed optimal decisions of each entity are consistent with the global optimal strategy.

[0123] The specific implementation method is as follows: 1. Multi-Agent Intelligent Agent Mapping and Action Space Division: Based on the division of business responsibilities for airport ground support, five distributed intelligent agents are constructed, namely wheel chock maintenance intelligent agent, boarding bridge operation intelligent agent, cabin support intelligent agent, passenger service intelligent agent, and pushback support intelligent agent. Each intelligent agent corresponds to the key support events and influencing factors under the responsibility of the aforementioned entity, and can only control intervention actions within its own scope of responsibility, forming an independent action space for each intelligent agent and avoiding cross-entity actions exceeding authority.

[0124] 2. Construction of the Causal Potential Function: Based on the quantification results of the causal effects of each intervention action output in step five above, a global causal potential function is constructed. The global maximum value of this potential function corresponds to the globally optimal time constraint strategy mentioned above. Simultaneously, the individual payoff function of each agent perfectly matches the global causal potential function, forming a causal potential game. Specifically, the global causal potential function is defined as: In the formula, N represents the total number of agents. The intervention action selected for the i-th agent. For action The corresponding average intervention effect, For action and The joint causal effect is calculated based on the aforementioned quantification results of the joint intervention interval effect; the individual payoff function for each agent is... ,in The action combination of other agents ensures that the game is a potential game, and its pure policy Nash equilibrium point is the maximum point of the global causal potential function, which is the aforementioned global optimal policy.

[0125] 3. Distributed Policy Learning and Cooperative Convergence: Based on the constructed causal potential game model, a distributed optimal response learning algorithm is designed. Each agent iteratively updates its optimal action policy based solely on its own causal effect information and the historical action information of other agents, without needing to obtain global state information. The specific iteration rule is as follows: In each iteration, the i-th agent, given that the actions of other agents are fixed, chooses the action that maximizes its own reward function. The process continues until all agents' actions cease to change, reaching Nash equilibrium. This iterative process is deeply integrated with the path integral control algorithm described above, using the combined actions of each agent after convergence as the core constraint of the initial behavioral policy set in step seven above, ensuring that the initial policy naturally meets the requirements of multi-agent collaboration.

[0126] 4. Cooperative conflict resolution mechanism: If multi-agent action conflicts occur during the iteration process (such as the causal effects of two agents' actions on the same key event canceling each other out), the core causal path of the conflict is identified based on the causal graph, and the cooperative conflict is resolved by adjusting the priority of the intervention interval of the conflicting actions. Specifically, the action with the larger absolute value of the total causal effect is retained, and the intervention interval of the other action is adjusted to ensure that the joint causal effect of the two actions is a positive gain and to avoid mutual cancellation.

[0127] 5. Integration with the closed loop of the entire process described above: The optimal action policies of each agent obtained through distributed learning are synchronously input into the path integral control algorithm in step eight above, serving as the initial policy and action search constraint for algorithm iteration; at the same time, the globally optimal policy output in step eight above will update the reward function and causal potential function of each agent in reverse, forming a two-way closed loop of distributed collaboration and global optimization.

[0128] This embodiment overcomes the inherent contradiction between centralized optimization and distributed implementation in existing technologies. It deeply integrates causal effect quantification with potential game theory to construct a multi-subject causal potential game model for civil aviation ground support. Mathematically, it ensures the consistency between the distributed optimal decision and the global optimal strategy of each support subject. This solves the problem of matching rights and responsibilities in the implementation of centralized strategies and avoids the global efficiency loss caused by distributed decision-making.

[0129] In some embodiments, this solution was piloted for three months at a hub airport in China with an annual passenger throughput exceeding 40 million. The pilot scope covered the entire process of ground support management for domestic narrow-body aircraft flights at the airport. During the pilot application, based on the time management requirements for the entire ground support process generated by this solution, the arrival time of airport ground staff and the operation time of each support link were pre-controlled and dynamically scheduled. During the pilot period, the on-time departure rate of the pilot flights at the airport steadily improved compared to the same period before the pilot, the average total time of ground support links was effectively reduced compared to the same period before the pilot, and the abnormal fluctuation range of support links was significantly reduced. The strategy maintained stable control effects under complex operating scenarios such as morning peak, evening peak, and extreme weather, verifying the practical application value and generalization ability of this solution.

Claims

1. A time-constrained optimization method for flight ground support events based on causal reasoning, characterized in that, Includes the following steps: Step 1: Obtain the time node dataset of the entire historical flight support process, then construct a hierarchical analysis system for flight ground support, determine the key events of the target layer support and the corresponding control factors of the influencing layer, design a unified standardized quantitative reference quantity, and obtain the standardized quantitative reference quantity dataset after preprocessing. Step 2: Combining the hierarchical analysis system of flight ground support, construct a set of prior constraints for civil aviation business and generate a causal structure feasible domain matrix; Step 3: Based on the standardized quantitative reference dataset, and with the civil aviation business prior constraint set and the causal structure feasible region matrix as constraints, the NOTEARS-MCP algorithm with adjacency matrix smoothing constraint is used to learn the sparse causal adjacency matrix and causal graph of flight ground support events. Step 4: Based on the standardized quantitative reference dataset and the sparse causal adjacency matrix, the standardized quantitative reference quantities of the factors affecting flight ground support are divided into interval heterogeneity. After compliance verification, the effective intervention interval set corresponding to each key event is generated. Step 5: Based on the standardized quantitative reference dataset, causal graph, and effective intervention interval set, the causal effect estimation algorithm based on entropy balance weighting is used to calculate the average intervention effect corresponding to each effective intervention interval and generate a causal intensity ranking table of intervention intervals. Step 6: Based on the sparse causal adjacency matrix, the average intervention effect and the causal intensity ranking table of the intervention interval, construct the Markov decision process model for flight ground support, and obtain the compressed state-action space set after state-action space compression. Step 7: Based on the average intervention effect, the causal intensity ranking table of the intervention interval, and the compressed state-action space set, construct an initial behavioral strategy set with dual constraints of causal intensity. Step 8: Based on the sparse causal adjacency matrix, average intervention effect, Markov decision process model, compressed state-action space set and initial behavior strategy set, the path integral control algorithm based on linearly solvable Markov decision process is used to solve for the global optimal time constraint strategy for the entire process of flight ground support. Step nine: Perform robust perturbation verification on the globally optimal time constraint strategy to obtain the final globally optimal time constraint strategy that passes the verification.

2. The time-constrained optimization method for flight ground support events based on causal reasoning according to claim 1, characterized in that, It also includes step ten, which transforms the final globally optimal time constraint strategy into a feasible time control requirement for the entire flight ground support process, by combining the hierarchical analysis system and standardized quantitative reference calculation rules of flight ground support.

3. The time-constrained optimization method for flight ground support events based on causal reasoning according to claim 1, characterized in that, The civil aviation business prior constraint set constructed in step two includes two types: prohibitive constraints and mandatory constraints. Prohibitive constraints are used to eliminate causal directions that do not conform to the business time sequence, while mandatory constraints are used to retain causal dependency edges that conform to the business logic. The causal structure feasible region matrix is ​​a binary matrix generated based on the two types of constraints, which is used to limit the search boundary of the causal structure.

4. The time-constrained optimization method for flight ground support events based on causal reasoning according to claim 1, characterized in that, The NOTEARS-MCP algorithm with adjacency matrix smoothing constraints in step three transforms the discrete search of the directed acyclic graph into a continuous optimization problem. It constructs the objective function with the data fitting goodness term, the MCP non-convex penalty term, the prior constraint term, and the acyclic smoothing constraint term, and solves iteratively through the ADMM optimization algorithm. Finally, the sparse causal adjacency matrix is ​​obtained after thresholding.

5. The time-constrained optimization method for flight ground support events based on causal reasoning according to claim 1, characterized in that, In step four, the heterogeneity partitioning of intervals is first completed using the equal frequency quantile partitioning method. Then, the interval boundaries are adjusted in combination with the time scale of civil aviation ground support operations. After compliance verification through sample size verification and covariate overlap verification, the effective intervention interval set is generated after removing the intervals that do not meet the requirements.

6. The time-constrained optimization method for flight ground support events based on causal reasoning according to claim 1, characterized in that, The causal effect estimation algorithm based on entropy balance weighting in step five aims to minimize the entropy loss of the control group sample weights and solves for the optimal weights under the constraint of complete balance of covariate moments between the intervention group and the control group. Based on the optimal weights, the mean difference in the execution time of the target key events between the intervention group and the weighted control group is calculated to obtain the average intervention effect of the corresponding intervention interval.

7. The time-constrained optimization method for flight ground support events based on causal reasoning according to claim 1, characterized in that, In the Markov decision process model for flight ground support constructed in step six, state-action space compression is based on the causal intensity ranking table to eliminate invalid actions with low impact intensity, thereby reducing the strategy search space.

8. The time-constrained optimization method for flight ground support events based on causal reasoning according to claim 1, characterized in that, The initial behavioral strategy set constructed in step seven with dual constraints on causal strength first calculates the initial causal weight of each action based on the average intervention effect, then corrects the initial weights by the state change benefits of the action execution, and obtains the selection probability of each action in each state after normalization, thus forming the initial behavioral strategy set.

9. The time-constrained optimization method for flight ground support events based on causal reasoning according to claim 1, characterized in that, The path integral control algorithm based on a linearly solvable Markov decision process in step eight transforms the nonlinear Bellman dynamic programming equation into a linearly solvable equation through exponential transformation. It then iteratively solves the state value function by combining the initial behavior policy set to obtain the optimal action probability distribution and finally extracts the global optimal time constraint policy for the entire process.

10. The time-constrained optimization method for flight ground support events based on causal reasoning according to claim 1, characterized in that, In step nine, the robustness perturbation check applies a fixed proportion of positive and negative perturbations to the average intervention effect corresponding to all actions in the globally optimal time-constrained strategy. Based on the results after the perturbation, the causal strength ranking is recalculated, and the change rate of the ranking before and after the perturbation is compared to determine whether the strategy passes the check. If the check fails, it is fed back to step four to adjust the interval partitioning granularity, and the corresponding steps are repeated to solve for the optimal strategy again.

Images