Aviation airport ground service intelligent scheduling and optimizing method based on deep learning

Abstract

The invention discloses an aviation airport ground service intelligent scheduling and optimization method based on deep learning, and the method comprises the following steps: S1, collecting task data, resource states and scheduling records, and constructing an input data set; s2, constructing a task and resource heterogeneous graph, and extracting a coupling representation; s3, constructing a priori knowledge constraint set, and embedding a state control module; s4, coding a service time window, fusing and inputting into a gating mechanism module; s5, key events are extracted, and sparse event enhancement representation is generated; s6, splicing various types of information to form a scheduling input sequence, and inputting the scheduling input sequence into the improved Hyena model to generate a scheduling sequence; and S7, generating a resource scheduling instruction according to the scheduling sequence, and completing task execution. According to the invention, intelligent scheduling and automatic optimization in a complex airport scene are realized.

Description

Technical Field

[0001] This invention relates to the field of transportation management technology, and in particular to a method for intelligent scheduling and optimization of airport ground services based on deep learning. Background Technology

[0002] Airport ground service scheduling is a core component of the modern civil aviation operation support system, encompassing a variety of tasks such as aircraft guidance, boarding bridge connection, baggage handling, refueling and replenishment, cleaning and preparation, and in-flight catering. These tasks are highly time-sensitive and require strict coordination of ground resources such as manpower and equipment. There are also close sequential constraints and dependencies between different tasks. In actual operation, airports need to dynamically adjust ground service scheduling plans based on flight schedules, resource availability, and unforeseen circumstances to meet multiple objectives of safety, punctuality, and efficiency.

[0003] Traditional scheduling methods often rely on rule bases and human experience, or use a combination of optimization algorithms such as heuristics, branch and bound, and integer programming for modeling. While these methods are effective in small to medium-sized scenarios, they often fail to meet the dual requirements of real-time performance and optimality when facing airport operation environments with dense resource conflicts and complex intertwined tasks. In addition, existing methods lack the ability to model complex state spaces and have limited utilization of historical scheduling data, task dependencies, and resource temporal constraints. The scheduling results are prone to problems such as bottleneck resource conflicts and time window mismatches.

[0004] In recent years, the application of deep learning technology in scheduling optimization has gradually emerged. Some studies have begun to explore the introduction of graph neural networks, sequence modeling networks, and attention mechanisms into ground service scheduling modeling to mine the implicit relationships between tasks and resources and predict scheduling sequences in an end-to-end manner. These methods have shown some potential in improving the level of scheduling automation, but there are still some shortcomings. On the one hand, existing methods generally represent tasks and resources as isomorphic graphs or single input sequences, failing to accurately capture the type interaction and coupling characteristics in heterogeneous structures. On the other hand, time window constraints and key sparse events are often simplified or ignored in the model, making it difficult for the model to maintain scheduling stability in time-intensive and event-driven scenarios. In addition, prior knowledge such as task order, resource constraints, and scheduling rules is difficult to embed into the modeling process, resulting in insufficient state space control capability of the model and easy occurrence of redundant paths and unreasonable scheduling.

[0005] In summary, existing technologies still have significant shortcomings in terms of modeling structure representation, timing control, event response, and prior constraints, and further breakthroughs and improvements are urgently needed in complex airport operation environments.

[0006] Therefore, how to provide intelligent scheduling and optimization methods for airport ground services based on deep learning is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0007] One objective of this invention is to propose an intelligent scheduling and optimization method for airport ground services based on deep learning. This invention comprehensively introduces structure-aware convolutional modeling, prior knowledge constraint mechanism, time window signal modeling and sparse event enhancement technology, and constructs an improved Hyena model that integrates task-resource heterogeneous graph structure and scheduling rule constraints. It also describes in detail a method for generating scheduling sequences for airport ground service scenarios, which has the advantages of strong structural modeling capabilities, reasonable task execution timing, accurate response to key events, and high scalability of scheduling strategies.

[0008] The intelligent scheduling and optimization method for airport ground services based on deep learning according to embodiments of the present invention includes the following steps:

[0009] S1. Collect ground service task data, ground resource status data and historical scheduling records of airports to construct the input dataset;

[0010] S2. Construct a heterogeneous graph structure of tasks and resources, and use structure-aware convolutional kernels to extract joint features between task nodes and resource nodes to generate a task-resource coupled representation.

[0011] S3. The task sequence, resource constraints, and scheduling rules are formed into a set of prior knowledge constraints, which are then embedded into the state control module of the improved Hyena model to limit the state space transition.

[0012] S4. Encode the start and end times of the service time window into a time window signal, fuse it with the task-resource coupling representation, and input it into the gating mechanism module of the improved Hyena model;

[0013] S5. Extract key event markers, construct a sparse event feature set, input it into the sparse event enhancement module, and generate a sparse event enhanced representation.

[0014] S6. The task-resource coupling representation, prior knowledge constraint set, time window signal and sparse event enhanced representation are spliced ​​together to form a scheduling input sequence, which is then input into the improved Hyena model to generate a ground service scheduling sequence.

[0015] S7. Generate ground resource scheduling instructions based on the ground service scheduling sequence to complete the scheduling and execution of airport ground service tasks.

[0016] Optionally, the input dataset includes task identifiers, service time windows, resource types, task dependencies, and key event markers.

[0017] Optionally, the improved Hyena model introduces four modules:

[0018] The structure-aware convolutional kernel module is used to receive the heterogeneous graph structure composed of tasks and resources, and extract the coupling features between task nodes and resource nodes through structure-aware convolution.

[0019] The state control module is used to embed prior knowledge constraint information formed by task order, resource constraints and scheduling rules, and to control the state transition path in the state space;

[0020] The gating mechanism module is used to receive the time window signal consisting of the start time and end time of the time window, and adjust the temporal arrangement of tasks in the sequence modeling process according to the time window signal;

[0021] The sparse event enhancement module is used to process key sparse event features in the input data and use these sparse event features as part of the model input for feature modeling.

[0022] Optionally, S2 specifically includes:

[0023] S21. Construct a heterogeneous graph structure G = (V) between service tasks and ground resources. T ∪V R ,E), where V T V represents the set of service task nodes. R Let E represent the set of ground resource nodes, and let E represent the set of edges connecting service task nodes and ground resource nodes.

[0024] S22. Generate initial embedding vectors for the service task nodes and ground resource nodes respectively. The embedding vector consists of task identifier, resource type, and basic attribute encoding, where d represents the embedding dimension;

[0025] S23. In the structure-aware convolutional kernel module, a structure attention mechanism is used to update the node embeddings:

[0026]

[0027] in, Let represent the representation vector of node i in the l-th layer. Let i represent the set of adjacent nodes connected to node i. Represents a linear transformation matrix. The structure-aware attention weights are calculated based on the differences between edge type encoding and node embedding, and σ(x) = max(0,x) represents the ReLU activation function.

[0028] S24. Pair the updated service task node representation with the ground resource node representation according to the actual scheduling association, and generate a task-resource coupled representation through feature concatenation and linear mapping. d′ represents the dimension of the coupling representation.

[0029] Optionally, S3 specifically includes:

[0030] S31. Extract task order information from the input dataset and construct a task order constraint set C according to the priority relationship between tasks. s ={(t i ,t j )}, where t i Indicates the task identifier, (t) i ,t j ) represents task t i Should be in task t j Complete before execution;

[0031] S32. Extract resource usage restriction information and construct a resource restriction set C based on the resource status within a specific time interval. r ={(r k ,β k )}, where r k Indicates resource type identifier, β k ∈{0,1} indicates whether the resource is available, with 0 indicating unavailable and 1 indicating available;

[0032] S33. Extract scheduling rule information from the input dataset and construct a scheduling rule set C. p ={p m}, where p m This represents the scheduling rule identifier according to the predefined rule set number, represented by a one-hot encoded vector of length l, where l is the total number of rule categories;

[0033] S34. Set the task order constraint set C s Resource constraint set C r and the scheduling rule set C p The combination forms a set of prior knowledge constraints C = [C s C r C p ];

[0034] S35. Input the set of prior knowledge constraints C into the state control module, and integrate it with the task-resource coupling representation within the state control module to define the transition structure between each state in the state space.

[0035] Optionally, S4 specifically includes:

[0036] S41. Extract the service time window information for each service task from the input dataset and record the start time of the service task. With deadline Where t start Indicates the earliest executable time of the service task, t endIndicates the latest time to complete the service task;

[0037] S42. Encode the service time window information into a time window signal vector. Where i represents the service task number;

[0038] S43, convert the time window signal vector w i Task-resource coupling representation corresponding to service tasks The vectors are concatenated to form a fused representation vector.

[0039] S44. Merge the representation vector f i Input the gating mechanism module of the improved Hyena model;

[0040] S45. In the gating mechanism module, based on the time window signal vector w i Calculate the time location weight of service task i Calculate the time and location weights:

[0041] α i =σ(W1·t start,i +W2·t end,i +b);

[0042] in, and For learnable scalar weight parameters, For bias terms, For the Sigmoid function;

[0043] S46. Weight the time location α i Used as a control signal, it adjusts the order of service tasks i in the input sequence in the improved Hyena model, and controls the activation time step of service tasks when performing scheduling sequence modeling in the improved Hyena model.

[0044] Optionally, S5 specifically includes:

[0045] S51. Extract key event markers from the input dataset, and identify the position index i and corresponding event type identifier of each key event in the scheduling input sequence.

[0046] S52. Calculate the sparsity weights of key events.

[0047]

[0048] in, Indicates event type e i The frequency of occurrence in the input dataset, t iIndicates the time position of a critical event on the scheduling timeline, t ref The center position of the scheduling time range is represented by λ1, T represents the total span of the scheduling time range, and λ1 and λ2 represent the weighting parameters of the frequency factor and the time factor. This represents the Sigmoid function, used to normalize the output range;

[0049] S53, Identify the event type e i With sparsity weight γ i Encoded as sparse event feature vectors Where k represents the total number of critical event types, onehot(e i ) indicates event type e i One-hot encoding representation;

[0050] S54. Transfer the sparse event feature vector s i The input is a sparse event augmentation module, which generates a sparse event augmentation representation through a structure including a linear transformation layer, a ReLU activation function, and attention weight scaling.

[0051] Optionally, S6 specifically includes:

[0052] S61. For each service task, represent the task-resource coupling. Prior knowledge constraint set representation Time window signal vector Enhanced representation of sparse events The vectors are concatenated to form the scheduling input vector. Where d = d1 + d2 + 2 + d3;

[0053] S62. Based on the start time in the time window signal, adjust the scheduling input vector x. i Arrange in ascending order to form the scheduling input sequence. Where n represents the number of service tasks;

[0054] S63. Input the scheduling input sequence X into the structure-aware convolutional kernel module of the improved Hyena model to generate an intermediate representation sequence. Where d′ represents the convolution feature dimension;

[0055] S64, Transform the intermediate representation sequence H conv The ground service scheduling sequence is generated by sequentially inputting the state control module, gating mechanism module, and sparse event enhancement module. Where y i Let m represent the scheduling output vector corresponding to service task i, and m represent the dimension of the scheduling instruction.

[0056] Optionally, S7 specifically includes:

[0057] S71, from ground service dispatch sequence Extract the task identifier ID corresponding to each service task. i Scheduling start time Scheduling end time Resource type identifier r i and scheduling priority value p i This constitutes the scheduling instruction tuple.

[0058] S72, regarding the scheduling instruction tuple d i Applying the scheduling constraint function φ(d) i The scheduling constraint function verifies whether the task dependencies, service time window consistency, and resource type matching conditions are met. If all conditions are met, then φ(d) is executed. i ) = 1, otherwise let φ(d) = 1. i ) = 0;

[0059] S73. Form a scheduling instruction set D = {d} by combining instruction tuples that satisfy the scheduling constraint function. i ∣φ(d i )=1}, and convert each scheduling instruction tuple into a control instruction vector. Form a control command sequence C = {c1, c2, ..., c k}, where k≤n, and q represents the parameter dimension of the control command;

[0060] S74. Input the control command sequence C into the airport ground service task scheduling system, and the scheduling system executes the ground resource scheduling commands, including issuing scheduling commands, updating resource status, and starting service task execution.

[0061] The beneficial effects of this invention are:

[0062] This invention constructs a heterogeneous graph structure that integrates service tasks and ground resources, and introduces a structure-aware convolutional kernel module. This enables the model to extract coupling features from the topological relationship between task nodes and resource nodes, thereby improving the structural expressiveness of scheduling modeling. This structure not only preserves the multidimensional attributes between tasks and resources, but also enhances the scheduling scheme's understanding of structural matching and resource adaptation.

[0063] During the scheduling and control process, a set of prior knowledge constraints consisting of task order, resource constraints, and scheduling rules is adopted and embedded in the state control module to limit the transition path of the state space. This effectively avoids invalid search and conflict path selection during the scheduling process and improves the model's compliance with the rule constraints.

[0064] To address the characteristic of service tasks having a defined time window, a time window signal encoding mechanism was designed, and the time window signal was input into the gating mechanism module to guide the model to automatically adjust the temporal arrangement of tasks in the scheduling sequence during the sequence modeling process, thereby improving the time consistency and window fit of the scheduling results.

[0065] Considering that there are infrequent but significant critical sparse events during airport operations, such as flight delays and equipment failures, this invention introduces a sparse event enhancement module. This module models sparse event features as learnable representations and incorporates them into scheduling decisions, which helps improve the model's ability to identify critical situations and its dynamic response capabilities.

[0066] Through the integration of the above-mentioned structured design and feature modeling mechanisms, the overall method can generate scheduling sequences with strong scheduling feasibility, reasonable task order, and efficient resource allocation in airport scenarios with high ground task complexity, dense resource conflicts, and strict scheduling constraints, without relying on explicit optimization solutions. Attached Figure Description

[0067] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0068] Figure 1 This is a flowchart of the intelligent scheduling and optimization method for airport ground services based on deep learning proposed in this invention;

[0069] Figure 2 This is a schematic diagram of the scheduling input feature modeling structure of the deep learning-based intelligent scheduling and optimization method for airport ground services proposed in this invention.

[0070] Figure 3 This is a schematic diagram of the improved Hyena model structure of the deep learning-based intelligent scheduling and optimization method for airport ground services proposed in this invention. Detailed Implementation

[0071] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0072] refer to Figure 1-3 A deep learning-based intelligent scheduling and optimization method for airport ground services includes the following steps:

[0073] S1. Collect ground service task data, ground resource status data and historical scheduling records of airports to construct the input dataset;

[0074] S2. Construct a heterogeneous graph structure of tasks and resources, and use structure-aware convolutional kernels to extract joint features between task nodes and resource nodes to generate a task-resource coupled representation.

[0075] S3. The task sequence, resource constraints, and scheduling rules are formed into a set of prior knowledge constraints, which are then embedded into the state control module of the improved Hyena model to limit the state space transition.

[0076] S4. Encode the start and end times of the service time window into a time window signal, fuse it with the task-resource coupling representation, and input it into the gating mechanism module of the improved Hyena model;

[0077] S5. Extract key event markers, construct a sparse event feature set, input it into the sparse event enhancement module, and generate a sparse event enhanced representation.

[0078] S6. The task-resource coupling representation, prior knowledge constraint set, time window signal and sparse event enhanced representation are spliced ​​together to form a scheduling input sequence, which is then input into the improved Hyena model to generate a ground service scheduling sequence.

[0079] S7. Generate ground resource scheduling instructions based on the ground service scheduling sequence to complete the scheduling and execution of airport ground service tasks.

[0080] This invention establishes an end-to-end deep learning scheduling method that integrates ground service task data, resource status data, and historical scheduling information into a unified modeling input. It also implements task-resource relationship modeling, timing constraint control, and critical event enhancement processing through structured steps, thereby improving the scheduling rationality and automated execution efficiency of airport ground service tasks.

[0081] In this embodiment, the input dataset includes task identifiers, service time windows, resource types, task dependencies, and key event markers.

[0082] This invention clearly defines the composition of the input dataset, including task identifiers, service time windows, resource types, task dependencies, and key event markers. This helps improve the standardization of the data structure and the consistency of the model input, ensuring that the scheduling model can accurately reflect business constraints and actual operating scenarios.

[0083] In this embodiment, the improved Hyena model introduces four modules:

[0084] The structure-aware convolutional kernel module is used to receive the heterogeneous graph structure composed of tasks and resources, and extract the coupling features between task nodes and resource nodes through structure-aware convolution.

[0085] The state control module is used to embed prior knowledge constraint information formed by task order, resource constraints and scheduling rules, and to control the state transition path in the state space;

[0086] The gating mechanism module is used to receive the time window signal consisting of the start time and end time of the time window, and adjust the temporal arrangement of tasks in the sequence modeling process according to the time window signal;

[0087] The sparse event enhancement module is used to process key sparse event features in the input data and use these sparse event features as part of the model input for feature modeling.

[0088] This invention introduces a structure-aware convolutional kernel, a state control module, a gating mechanism module, and a sparse event enhancement module on the basis of the Hyena model, so as to achieve collaborative processing of heterogeneous graph structure modeling, prior knowledge constraint control, time window driven modeling, and key event reinforcement, thereby enhancing the model's expressive ability and scheduling generation stability in complex scheduling scenarios.

[0089] In this embodiment, S2 specifically includes:

[0090] S21. Construct a heterogeneous graph structure G = (V) between service tasks and ground resources. T ∪V R ,E), where V T V represents the set of service task nodes. R Let E represent the set of ground resource nodes, and let E represent the set of edges connecting service task nodes and ground resource nodes.

[0091] S22. Generate initial embedding vectors for the service task nodes and ground resource nodes respectively. The embedding vector consists of task identifier, resource type, and basic attribute encoding, where d represents the embedding dimension;

[0092] S23. In the structure-aware convolutional kernel module, a structure attention mechanism is used to update the node embeddings:

[0093]

[0094] in, Let represent the representation vector of node i in the l-th layer. Let i represent the set of adjacent nodes connected to node i. Represents a linear transformation matrix. The structure-aware attention weights are calculated based on the differences between edge type encoding and node embedding, and σ(x) = max(0,x) represents the ReLU activation function.

[0095] S24. Pair the updated service task node representation with the ground resource node representation according to the actual scheduling association, and generate a task-resource coupled representation through feature concatenation and linear mapping. d′ represents the dimension of the coupling representation.

[0096] Based on the construction of a heterogeneous graph structure, this invention uses a structure-aware convolution method to extract coupled features between service task nodes and ground resource nodes, enabling the scheduling model to capture the topological relationships and attribute dependencies between tasks and resources, thereby improving the structural expression accuracy of the scheduling input representation.

[0097] In this embodiment, S3 specifically includes:

[0098] S31. Extract task order information from the input dataset and construct a task order constraint set C according to the priority relationship between tasks. s ={(t i ,t j )}, where t i Indicates the task identifier, (t) i ,t j ) represents task t i Should be in task t j Complete before execution;

[0099] S32. Extract resource usage restriction information and construct a resource restriction set C based on the resource status within a specific time interval. r ={(r k ,β k )}, where r k Indicates resource type identifier, β k ∈{0,1} indicates whether the resource is available, with 0 indicating unavailable and 1 indicating available;

[0100] S33. Extract scheduling rule information from the input dataset and construct a scheduling rule set C. p ={p m}, where p m This represents the scheduling rule identifier according to the predefined rule set number, represented by a one-hot encoded vector of length l, where l is the total number of rule categories;

[0101] S34. Set the task order constraint set C s Resource constraint set C r and the scheduling rule set C p The combination forms a set of prior knowledge constraints C = [C s C r C p ];

[0102] S35. Input the set of prior knowledge constraints C into the state control module, and integrate it with the task-resource coupling representation within the state control module to define the transition structure between each state in the state space.

[0103] This invention extracts task order, resource constraints, and scheduling rules from input data, constructs a set of prior knowledge constraints, and defines a state transition structure in the state control module. This enables the model to have rule-guided capabilities during sequence generation, reduces the risk of generating illegal paths, and improves the executability of scheduling.

[0104] In this embodiment, S4 specifically includes:

[0105] S41. Extract the service time window information for each service task from the input dataset and record the start time of the service task. With deadline Where t start Indicates the earliest executable time of the service task, t end Indicates the latest time to complete the service task;

[0106] S42. Encode the service time window information into a time window signal vector. Where i represents the service task number;

[0107] S43, convert the time window signal vector w i Task-resource coupling representation corresponding to service tasks The vectors are concatenated to form a fused representation vector.

[0108] S44. Merge the representation vector f i Input the gating mechanism module of the improved Hyena model;

[0109] S45. In the gating mechanism module, based on the time window signal vector w i Calculate the time location weight of service task i Calculate the time and location weights:

[0110] α i =σ(W1·t start,i +W2·t end,i +b);

[0111] in, and These are learnable scalar weight parameters. For bias terms, For the Sigmoid function;

[0112] S46. Weight the time location α iUsed as a control signal, it adjusts the order of service tasks i in the input sequence in the improved Hyena model, and controls the activation time step of service tasks when performing scheduling sequence modeling in the improved Hyena model.

[0113] This invention enhances the model's sensitivity to time constraints by encoding the service time window into a time window signal and embedding it into a gating mechanism module. It combines time position weights to dynamically sort and control the input tasks, making the scheduling sequence more consistent with the actual available time of the task and reducing scheduling conflicts.

[0114] In this embodiment, S5 specifically includes:

[0115] S51. Extract key event markers from the input dataset, and identify the position index i and corresponding event type identifier of each key event in the scheduling input sequence.

[0116] S52. Calculate the sparsity weights of key events.

[0117]

[0118] in, Indicates event type e i The frequency of occurrence in the input dataset, t i Indicates the time position of a critical event on the scheduling timeline, t ref The center position of the scheduling time range is represented by λ1, T represents the total span of the scheduling time range, and λ1 and λ2 represent the weighting parameters of the frequency factor and the time factor. This represents the Sigmoid function, used to normalize the output range;

[0119] S53, Identify the event type e i With sparsity weight γ i Encoded as sparse event feature vectors Where k represents the total number of critical event types, onehot(e i ) indicates event type e i One-hot encoding representation;

[0120] S54. Transfer the sparse event feature vector s i The input is a sparse event augmentation module, which generates a sparse event augmentation representation through a structure including a linear transformation layer, a ReLU activation function, and attention weight scaling.

[0121] This invention performs sparsity modeling and event feature enhancement processing on key sparse events. By introducing an attention mechanism and nonlinear mapping to generate enhanced representations of sparse events, it strengthens the model's ability to perceive low-frequency, high-importance events and improves the robustness of scheduling schemes in handling special cases.

[0122] In this embodiment, S6 specifically includes:

[0123] S61. For each service task, represent the task-resource coupling. Prior knowledge constraint set representation Time window signal vector Enhanced representation of sparse events The vectors are concatenated to form the scheduling input vector. Where d = d1 + d2 + 2 + d3;

[0124] S62. Based on the start time in the time window signal, adjust the scheduling input vector x. i Arrange in ascending order to form the scheduling input sequence. Where n represents the number of service tasks;

[0125] S63. Input the scheduling input sequence X into the structure-aware convolutional kernel module of the improved Hyena model to generate an intermediate representation sequence. Where d′ represents the convolution feature dimension;

[0126] S64, Transform the intermediate representation sequence H conv The ground service scheduling sequence is generated by sequentially inputting the state control module, gating mechanism module, and sparse event enhancement module. Where y i Let m represent the scheduling output vector corresponding to service task i, and m represent the dimension of the scheduling instruction.

[0127] This invention constructs a scheduling input sequence by concatenating task-resource coupled representation, prior knowledge constraint set, time window signal and sparse event enhanced representation, and processes the input information sequentially through a structured model module, thereby improving the rationality of information flow and feature fusion within the model and providing a more stable semantic foundation for subsequent scheduling generation.

[0128] In this embodiment, S7 specifically includes:

[0129] S71, from ground service dispatch sequence Extract the task identifier ID corresponding to each service task. i Scheduling start time Scheduling end time Resource type identifier r i and scheduling priority value p i This constitutes the scheduling instruction tuple.

[0130] S72, regarding the scheduling instruction tuple d i Applying the scheduling constraint function φ(d) i The scheduling constraint function verifies whether the task dependencies, service time window consistency, and resource type matching conditions are met. If all conditions are met, then φ(d) is executed. i ) = 1, otherwise let φ(d) = 1. i ) = 0;

[0131] S73. Form a scheduling instruction set D = {d} by combining instruction tuples that satisfy the scheduling constraint function. i ∣φ(d i )=1}, and convert each scheduling instruction tuple into a control instruction vector. Form a control command sequence C = {c1, c2, ..., c k}, where k≤n, and q represents the parameter dimension of the control command;

[0132] S74. Input the control command sequence C into the airport ground service task scheduling system, and the scheduling system executes the ground resource scheduling commands, including issuing scheduling commands, updating resource status, and starting service task execution.

[0133] This invention verifies and filters the instructions generated from the ground service scheduling sequence based on the scheduling constraint function, and converts feasible instructions into a sequence of control instructions for the scheduling system to execute, ensuring that the final scheduling execution meets the requirements of task dependency, time window consistency and resource matching, thereby improving the practical feasibility of the scheduling scheme.

[0134] Example 1:

[0135] To verify the feasibility of this invention in practice, it was applied to the summer transport operation management of a large international hub airport in East China in May 2025. The airport has an annual passenger throughput of over 80 million and an average daily flight takeoff and landing volume of over 1,300. Ground service tasks are intensive and resources are limited. The scheduling process is complex and anomalies occur frequently. The original scheduling system relies on rule bases and manual intervention, which is no longer able to meet the dual requirements of real-time performance and stability.

[0136] In actual deployment, the scheduling system first connects to the airport scheduling platform, resource control system, and A-CDM collaborative platform to collect service task lists, ground resource status, task dependency chains, time window constraints, and historical scheduling records. It then constructs a heterogeneous graph structure of tasks and resources, extracts the coupling features between task nodes and resource nodes through structure-aware convolutional kernels, and combines task sequence, resource rules, and scheduling constraints into a set of prior knowledge constraints. This set is then embedded in the state control module to restrict state space transition paths. Service time window start and end information is encoded into time window signals and input into the gating mechanism module to regulate the modeling timing of tasks. Key sparse events such as delay alarms, resource conflicts, and transfer flights are marked and processed into sparse event feature sets and input into the sparse event enhancement module for modeling.

[0137] Finally, the sequence is spliced ​​to generate a scheduling input sequence, which is then input into the improved Hyena model for inference. The output ground service scheduling sequence is then parsed into ground instructions by the scheduling system and automatically sent to various ground service equipment terminals to achieve closed-loop execution of task assignment and resource control.

[0138] The test period was selected from 08:00 to 14:00 on August 15, 2024. The system processed a total of 542 tasks, covering baggage handling, boarding bridge operation, flight guidance, and catering, involving 218 types of resources and equipment. The performance of the system of this invention was compared with that of the original scheduling system under the same conditions. The comparison results are shown in Table 1.

[0139] Table 1. Performance Comparison of Airport Dispatch Systems

[0140] index Existing scheduling system This invention system Average scheduling response delay (seconds) 18.3 4.6 Task execution overlap rate (%) 23.1 8.7 Default rate within the time window (%) 17.4 2.9 Resource scheduling conflict rate (per 100 times) 9.2 1.5 Success rate of handling abnormal tasks (%) 84.7 98.1 Daily number of automatically scheduled tasks 800 1600

[0141] As shown in Table 1, the system of this invention achieves significant improvements in several core indicators, including scheduling response time, task conflict rate, time window matching degree, abnormal event handling capability, and system scheduling throughput. Especially when dealing with sudden flight delays or resource chain failures, the system can quickly correct scheduling paths by combining prior knowledge and event enhancement mechanisms, ensuring operational continuity and timeliness.

[0142] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for intelligent scheduling and optimization of airport ground services based on deep learning, characterized in that, Includes the following steps: S1. Collect ground service task data, ground resource status data and historical scheduling records of airports to construct the input dataset; S2. Construct a heterogeneous graph structure of tasks and resources, and use structure-aware convolutional kernels to extract joint features between task nodes and resource nodes to generate a task-resource coupled representation. S3. The task sequence, resource constraints, and scheduling rules are formed into a set of prior knowledge constraints, which are then embedded into the state control module of the improved Hyena model to limit the state space transition. S4. Encode the start and end times of the service time window into a time window signal, fuse it with the task-resource coupling representation, and input it into the gating mechanism module of the improved Hyena model; S5. Extract key event markers, construct a sparse event feature set, input it into the sparse event enhancement module, and generate a sparse event enhanced representation. S6. The task-resource coupling representation, prior knowledge constraint set, time window signal and sparse event enhanced representation are spliced ​​together to form a scheduling input sequence, which is then input into the improved Hyena model to generate a ground service scheduling sequence. S7. Generate ground resource scheduling instructions based on the ground service scheduling sequence to complete the scheduling and execution of airport ground service tasks.

2. The method for intelligent scheduling and optimization of airport ground services based on deep learning according to claim 1, characterized in that, The input dataset includes task identifiers, service time windows, resource types, task dependencies, and key event markers.

3. The method for intelligent scheduling and optimization of airport ground services based on deep learning according to claim 1, characterized in that, The improved Hyena model introduces four modules: The structure-aware convolutional kernel module is used to receive the heterogeneous graph structure composed of tasks and resources, and extract the coupling features between task nodes and resource nodes through structure-aware convolution. The state control module is used to embed prior knowledge constraint information formed by task order, resource constraints and scheduling rules, and to control the state transition path in the state space; The gating mechanism module is used to receive the time window signal consisting of the start time and end time of the time window, and adjust the temporal arrangement of tasks in the sequence modeling process according to the time window signal; The sparse event enhancement module is used to process key sparse event features in the input data and use these sparse event features as part of the model input for feature modeling.

4. The method for intelligent scheduling and optimization of airport ground services based on deep learning according to claim 1, characterized in that, S2 specifically includes: S21. Construct a heterogeneous graph structure G = (V) between service tasks and ground resources. T ∪V R ,E), where V T V represents the set of service task nodes. R Let E represent the set of ground resource nodes, and let E represent the set of edges connecting service task nodes and ground resource nodes. S22. Generate initial embedding vectors for the service task nodes and ground resource nodes respectively. The embedding vector consists of task identifier, resource type, and basic attribute encoding, where d represents the embedding dimension; S23. In the structure-aware convolutional kernel module, a structure attention mechanism is used to update the node embeddings: in, Let represent the representation vector of node i in the l-th layer. Let i represent the set of adjacent nodes connected to node i. Represents a linear transformation matrix. The structure-aware attention weights are calculated based on the differences between edge type encoding and node embedding, and σ(x) = max(0,x) represents the ReLU activation function. S24. Pair the updated service task node representation with the ground resource node representation according to the actual scheduling association, and generate a task-resource coupled representation through feature concatenation and linear mapping. d′ represents the dimension of the coupling representation.

5. The method for intelligent scheduling and optimization of airport ground services based on deep learning according to claim 1, characterized in that, S3 specifically includes: S31. Extract task order information from the input dataset and construct a task order constraint set C according to the priority relationship between tasks. s ={(t i ,t j )}, where t i Indicates the task identifier, (t) i ,t j ) represents task t i Should be in task t j Complete before execution; S32. Extract resource usage restriction information and construct a resource restriction set C based on the resource status within a specific time interval. r ={(r k ,β k )}, where r k Indicates resource type identifier, β k ∈{0,1} indicates whether the resource is available, with 0 indicating unavailable and 1 indicating available; S33. Extract scheduling rule information from the input dataset and construct a scheduling rule set C. p ={p m }, where p m This represents the scheduling rule identifier according to the predefined rule set number, represented by a one-hot encoded vector of length l, where l is the total number of rule categories; S34. Set the task order constraint set C s Resource constraint set C r and the scheduling rule set C p The combination forms a set of prior knowledge constraints C = [C s C r C p ]; S35. Input the set of prior knowledge constraints C into the state control module, and integrate it with the task-resource coupling representation within the state control module to define the transition structure between each state in the state space.

6. The method for intelligent scheduling and optimization of airport ground services based on deep learning according to claim 1, characterized in that, S4 specifically includes: S41. Extract the service time window information for each service task from the input dataset and record the start time of the service task. With deadline Where t start Indicates the earliest executable time of the service task, t end Indicates the latest time to complete the service task; S42. Encode the service time window information into a time window signal vector. Where i represents the service task number; S43, convert the time window signal vector w i Task-resource coupling representation corresponding to service tasks The vectors are concatenated to form a fused representation vector. S44. Merge the representation vector f i Input the gating mechanism module of the improved Hyena model; S45. In the gating mechanism module, based on the time window signal vector w i Calculate the time location weight of service task i Calculate the time and location weights: a i =σ(W1·t start,i +W2·t end,i +b); in, and For learnable scalar weight parameters, For bias terms, For the Sigmoid function; S46. Weight the time location α i Used as a control signal, it adjusts the order of service tasks i in the input sequence in the improved Hyena model, and controls the activation time step of service tasks when performing scheduling sequence modeling in the improved Hyena model.

7. The method for intelligent scheduling and optimization of airport ground services based on deep learning according to claim 1, characterized in that, S5 specifically includes: S51. Extract key event markers from the input dataset, and identify the position index i and corresponding event type identifier of each key event in the scheduling input sequence. S52. Calculate the sparsity weights of key events. in, Indicates event type e i The frequency of occurrence in the input dataset, t i Indicates the time position of a critical event on the scheduling timeline, t ref The center position of the scheduling time range is represented by λ1, T represents the total span of the scheduling time range, and λ1 and λ2 represent the weighting parameters of the frequency factor and the time factor. This represents the Sigmoid function, used to normalize the output range; S53, Identify the event type e i With sparsity weight γ i Encoded as sparse event feature vectors Where k represents the total number of critical event types, onehot(e i ) indicates event type e i One-hot encoding representation; S54. Transfer the sparse event feature vector s i The input is a sparse event augmentation module, which generates a sparse event augmentation representation through a structure including a linear transformation layer, a ReLU activation function, and attention weight scaling.

8. The intelligent scheduling and optimization method for airport ground services based on deep learning according to claim 1, characterized in that, S6 specifically includes: S61. For each service task, represent the task-resource coupling. Prior knowledge constraint set representation Time window signal vector Enhanced representation of sparse events The vectors are concatenated to form the scheduling input vector. Where d = d1 + d2 + 2 + d3; S62. Based on the start time in the time window signal, adjust the scheduling input vector x. i Arrange in ascending order to form the scheduling input sequence. Where n represents the number of service tasks; S63. Input the scheduling input sequence X into the structure-aware convolutional kernel module of the improved Hyena model to generate an intermediate representation sequence. Where d′ represents the convolution feature dimension; S64, Transform the intermediate representation sequence H conv The ground service scheduling sequence is generated by sequentially inputting the state control module, gating mechanism module, and sparse event enhancement module. Where y i Let m represent the scheduling output vector corresponding to service task i, and m represent the dimension of the scheduling instruction.

9. The method for intelligent scheduling and optimization of airport ground services based on deep learning according to claim 1, characterized in that, Specifically, S7 includes: S71, from ground service dispatch sequence Extract the task identifier ID corresponding to each service task. i Scheduling start time Scheduling end time Resource type identifier r i and scheduling priority value p i This constitutes the scheduling instruction tuple. S72, regarding the scheduling instruction tuple d i Applying the scheduling constraint function φ(d) i The scheduling constraint function verifies whether the task dependencies, service time window consistency, and resource type matching conditions are met. If all conditions are met, then φ(d) is executed. i ) = 1, otherwise let φ(d) = 1. i ) = 0; S73. Form a scheduling instruction set D = {d} by combining instruction tuples that satisfy the scheduling constraint function. i ∣φ(d i )=1}, and convert each scheduling instruction tuple into a control instruction vector. Form a control command sequence C = {c1, c2, ..., c k }, where k≤n, and q represents the parameter dimension of the control command; S74. Input the control command sequence C into the airport ground service task scheduling system, and the scheduling system executes the ground resource scheduling commands, including issuing scheduling commands, updating resource status, and starting service task execution.

Images