Aircraft maintenance stand scheduling strategy determination method and device

Abstract

The invention provides an aircraft maintenance stand scheduling strategy determination method and device, and relates to the technical field of data processing, and the method comprises the steps: constructing a target loss function based on the number of movement times of aircrafts, the average maintenance stand utilization rate and the displacement of all aircrafts, taking an original feature vector and a first heterogeneous graph model as input, and carrying out the calculation of the target loss function; and training the maintenance stand scheduling prediction network by taking minimization of the target loss function as a target to obtain a target scheduling network, and obtaining an aircraft maintenance stand scheduling strategy through the target scheduling network, the input feature vector and the second heterogeneous graph model. The target scheduling network can meet the time constraint of the maintenance project and the space constraint of the maintenance machine position at the same time, and the solution quality and the solution efficiency under the complex maintenance machine position scheduling scene can be improved.

Description

Technical Field

[0001] The present application relates to the field of data processing technology, and in particular to a method and device for determining an aircraft maintenance stand scheduling strategy. Background Art

[0002] As the aircraft fleet expands, the booming demand for maintenance is leading to an increasing shortage of maintenance bay resources at aviation maintenance companies. The diversity of maintenance tasks, the complexity of the maintenance environment, and the high workload and high risk of maintenance work all place higher demands on the allocation and optimization of maintenance bay resources. Aircraft maintenance, repair, and overhaul (MRO) activities are crucial to the civil aviation industry. Aircraft maintenance is categorized as line maintenance and hangar maintenance based on the location of maintenance. Line maintenance refers to maintenance activities performed while the aircraft awaiting maintenance is parked on the jet bridge or the apron, including walk-around inspections and oiling. All other maintenance activities are classified as hangar maintenance and must be performed in a maintenance hangar. However, with the increase in aircraft awaiting maintenance and the number of maintenance tasks, the poor resource utilization of manual scheduling is severely restricting the full utilization of maintenance capacity. A more core issue is the continued increase in production safety pressures. Summary of the Invention

[0003] The present application provides a method and apparatus for determining an aircraft maintenance stand scheduling strategy. By constructing a target loss function based on the number of aircraft movements, the average maintenance stand utilization rate, and the displacement of all aircraft, and using the maintenance stand-maintenance stand arc feature vector as a model input, the target scheduling network can simultaneously meet the time constraints of maintenance projects and the spatial constraints of maintenance stands, ensuring the safety of maintenance stand scheduling strategies, and facilitating improved solution quality and efficiency in complex maintenance stand scheduling scenarios.

[0004] This application provides a method for determining an aircraft maintenance stand scheduling strategy, comprising: Obtaining original feature vectors, the original feature vectors including multiple maintenance item feature vectors, multiple maintenance bay feature vectors, a maintenance item-maintenance bay arc feature vector, and a maintenance bay-maintenance bay arc feature vector; the maintenance item-maintenance bay arc feature vector including the maintenance time for each maintenance item at each available maintenance bay; and the maintenance bay-maintenance bay arc feature vector including the distance between each maintenance bay and its adjacent maintenance bay; Constructing a first heterogeneous graph model based on the original feature vector; the first heterogeneous graph model includes maintenance project nodes, maintenance position nodes, connection arcs between each maintenance project and maintenance position, and connection arcs between each maintenance position; Construct a target loss function based on the number of aircraft moves, average maintenance bay utilization, and the displacement of all aircraft; Based on the original feature vector and the first heterogeneous graph model, training a maintenance aircraft bay scheduling prediction network with the goal of minimizing a target loss function to obtain a target scheduling network; Obtaining an input feature vector, the input feature vector including a plurality of maintenance item feature vectors of the aircraft to be maintained, a plurality of maintenance stand feature vectors, a feature vector of a maintenance item-maintenance stand arc, and a feature vector of a maintenance stand-maintenance stand arc; constructing a second heterogeneous graph model based on the input feature vector; The input feature vector and the second heterogeneous graph model are input into the target scheduling network to obtain an aircraft maintenance stand scheduling strategy determination strategy, so as to schedule the aircraft to be maintained based on the aircraft maintenance stand scheduling strategy determination strategy; the target scheduling network is used to obtain an output aircraft maintenance stand scheduling strategy based on the input feature vector and the second heterogeneous graph model.

[0005] To achieve the above-mentioned and other related objectives, the present application provides an aircraft maintenance stand scheduling strategy determination device, comprising: A first data acquisition module is configured to acquire original feature vectors, wherein the original feature vectors include a plurality of maintenance item feature vectors, a plurality of maintenance bay feature vectors, a maintenance item-maintenance bay arc feature vector, and a maintenance bay-maintenance bay arc feature vector. The maintenance item-maintenance bay arc feature vector includes the maintenance time for each maintenance item at each available maintenance bay. The maintenance bay-maintenance bay arc feature vector includes the distance between each maintenance bay and its adjacent maintenance bay. A first processing module is configured to construct a first heterogeneous graph model based on the original feature vector; the first heterogeneous graph model includes maintenance project nodes, maintenance station nodes, connection arcs between each maintenance project and maintenance station, and connection arcs between each maintenance station; A function building module is used to build a target loss function based on the number of aircraft movements, the average maintenance bay utilization rate, and the displacement of all aircraft; A model training module, configured to train a maintenance bay scheduling prediction network based on the original feature vector and the first heterogeneous graph model with the goal of minimizing a target loss function to obtain a target scheduling network; a second data acquisition module, configured to acquire an input feature vector, wherein the input feature vector includes a plurality of maintenance item feature vectors of the aircraft to be maintained, a plurality of maintenance stand feature vectors, a feature vector of a maintenance item-maintenance stand arc, and a feature vector of a maintenance stand-maintenance stand arc; A second processing module, configured to construct a second heterogeneous graph model based on the input feature vector; A strategy acquisition module is configured to input the input feature vector and the second heterogeneous graph model into the target scheduling network to obtain an aircraft maintenance stand scheduling strategy determination strategy, so as to schedule the aircraft to be maintained based on the aircraft maintenance stand scheduling strategy determination strategy; the target scheduling network is configured to obtain an output aircraft maintenance stand scheduling strategy determination strategy based on the input feature vector and the second heterogeneous graph model.

[0006] As described above, the present application provides a method and apparatus for determining an aircraft maintenance stand scheduling strategy, which has the following beneficial effects: This application discloses a method for determining an aircraft maintenance stand scheduling strategy. This method constructs a target loss function based on the number of aircraft moves, average maintenance stand utilization, and the displacement of all aircraft. The method uses the original feature vector and a first heterogeneous graph model as input, and trains a maintenance stand scheduling prediction network to minimize the target loss function, resulting in a target scheduling network. The aircraft maintenance stand scheduling strategy is then derived using the target scheduling network, the input feature vector, and the second heterogeneous graph model. The target scheduling network can simultaneously satisfy the time constraints of maintenance projects and the spatial constraints of maintenance stands, thereby improving the solution quality and efficiency in complex maintenance stand scheduling scenarios.

[0007] It should be understood that the foregoing general description and the following detailed description are exemplary and explanatory only and are not restrictive of the present application. BRIEF DESCRIPTION OF THE DRAWINGS

[0008] The accompanying drawings are incorporated into and constitute a part of the specification, illustrating embodiments consistent with the present application and, together with the specification, serving to explain the principles of the present application. It is obvious that the drawings described below are merely some embodiments of the present application, and a person of ordinary skill in the art can derive other drawings based on these drawings without inventive effort. In the drawings: Figure 1 is a flow chart of a method for determining an aircraft maintenance stand scheduling strategy, shown in an exemplary embodiment of the present application; Figure 2 is a schematic diagram of a heterogeneous graph model shown in an exemplary embodiment of the present application; Figure 3 1 is an architectural diagram of a maintenance bay scheduling prediction network according to an exemplary embodiment of the present application; Figure 4 It is a structural block diagram of an aircraft maintenance stand scheduling strategy determination device shown in an exemplary embodiment of the present application. DETAILED DESCRIPTION

[0009] The following will describe the embodiments of the present application with reference to the accompanying drawings and preferred embodiments. Those skilled in the art can easily understand the other advantages and effects of the present application from the contents disclosed in this specification. The present application can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed in various ways based on different viewpoints and applications without departing from the spirit of the present application. It should be understood that the preferred embodiments are only for the purpose of illustrating the present application and are not intended to limit the scope of protection of the present application.

[0010] It should be noted that the illustrations provided in the following embodiments are only schematic illustrations of the basic concept of the present application. Therefore, the illustrations only show components related to the present application and are not drawn according to the number, shape and size of components in actual implementation. In actual implementation, the type, quantity and proportion of each component can be changed at will, and the component layout type may also be more complicated.

[0011] In the following description, a large number of details are discussed to provide a more thorough explanation of the embodiments of the present application. However, it is obvious to those skilled in the art that the embodiments of the present application can be implemented without these specific details. In other embodiments, well-known structures and devices are shown in the form of block diagrams rather than in detail to avoid making the embodiments of the present application difficult to understand.

[0012] See also Figure 1 , Figure 1 FIG1 is a flow chart of a method for determining an aircraft maintenance stand scheduling strategy according to an exemplary embodiment of the present application. Figure 1 It can be seen that the method for determining the aircraft maintenance stand scheduling strategy may include: Step S110: Obtain the original feature vector.

[0013] Among them, the original feature vector includes multiple maintenance item feature vectors, multiple maintenance stand feature vectors, a maintenance item-maintenance stand arc feature vector, and a maintenance stand-maintenance stand arc feature vector; the maintenance item-maintenance stand arc feature vector includes the maintenance time of each maintenance item at each available maintenance stand; the maintenance stand-maintenance stand arc feature vector includes the distance between each maintenance stand and its adjacent maintenance stand.

[0014] In one embodiment of the present application, data can be determined based on an aircraft maintenance bay scheduling strategy to obtain an original feature vector, so as to provide data input for a subsequent node embedding process and a training process.

[0015] Each maintenance item feature vector includes the shortest maintenance time across all maintenance bays, the average maintenance time across all maintenance bays, the maintenance completion time across all maintenance bays, the scheduling status, the lower bound of the estimated completion time, the number of remaining maintenance items, the average maintenance time for unscheduled maintenance items, the proportion of bays performing the maintenance item, the waiting time, the remaining maintenance time, and a three-dimensional non-negative weight vector used during the training phase. For any maintenance item, not all maintenance bays have the capacity to handle it. Therefore, the number of available bays and the proportion of bays performing the maintenance item can be determined based on the maintenance capacity of each bay. If a maintenance item can be performed at any of two or more maintenance bays, the shortest maintenance time across all bays, the average maintenance time across all bays, the maintenance completion time across all bays, the average maintenance time for unscheduled maintenance items, and the lower bound of the estimated completion time can be determined. The number of remaining maintenance items can be determined based on the number of completed maintenance items for the aircraft to be repaired.

[0016] Each maintenance bay feature vector includes: the shortest maintenance time among all maintenance items, the average maintenance time of the maintenance items that can be handled, the number of maintenance items that can be handled but not yet scheduled, the idle time, the waiting time, the scheduling status, the remaining maintenance time, and a three-dimensional non-negative weight vector with the same settings as the maintenance item feature vector. The scheduling status can include both unscheduled and scheduled. In some embodiments, the unscheduled state can be represented by the number 0, and the scheduled state can be represented by the number 1.

[0017] It should be noted that during the network model training phase, the original feature vectors used for training are derived from data determined by the aircraft maintenance stand scheduling policy. Therefore, during the prediction phase, the start time of each maintenance item is known. However, the start time of the next maintenance item must be estimated. Therefore, when using the trained maintenance stand scheduling prediction network for prediction, the start time of each maintenance item is an estimated value. For example, this estimated value estimates the start time of the maintenance item based on the estimated start time and maintenance time of the previous maintenance item.

[0018] Step S120: construct a first heterogeneous graph model based on the original feature vector.

[0019] The first heterogeneous graph model includes maintenance project nodes, maintenance position nodes, connection arcs between each maintenance project and maintenance position, and connection arcs between each maintenance position.

[0020] In one embodiment of the present application, a heterogeneous graph model may be constructed based on the input feature vector. Both the first heterogeneous graph model and the second heterogeneous graph model may be referred to as heterogeneous graph models.

[0021] It should be noted that a maintenance bay is not only related to the currently processed maintenance project, but also to its adjacent bays. Furthermore, a set of arcs representing the distances between adjacent bays establishes spatial constraints through dynamic maintenance bays. Each undirected arc connects a bay node to its adjacent bay nodes, using the distance between bays as a feature to represent the spatial constraints.

[0022] For example, please refer to Figure 2 , maintenance project nodes include 、 、 、 、 、 、 、 、 、 and Maintenance stand nodes include 、 、 and Multiple maintenance items must be combined in a specific order. Each maintenance project can be processed on multiple maintenance bays with maintenance capabilities, but can only be processed on one of the idle and matching maintenance bays. The heterogeneous graph model includes two virtual nodes: all maintenance projects and all maintenance bays. Including all maintenance bays. A set of connected arcs is formed by connecting multiple directed paths from the start node to the end node. , represents the maintenance item sequence of each aircraft. Each undirected arc Connect the maintenance project node to the maintenance bay node that matches it. The corresponding processing time is In addition, a set of arcs representing the distances between adjacent maintenance bays By dynamically maintaining the aircraft position, the spatial constraints are established. Each undirected arc The maintenance stand node is connected to its adjacent maintenance stand nodes, and the distance between the maintenance stands is used as a feature to represent the spatial constraint.

[0023] Step S130 : constructing a target loss function based on the number of aircraft movements, the average maintenance bay utilization rate, and the displacement of all aircraft.

[0024] In one embodiment of the present application, the process of constructing a target loss function based on the number of aircraft movements, the average maintenance stand utilization rate, and the displacement of all aircraft may include: constructing a value function loss function based on the number of aircraft movements, the average maintenance stand utilization rate, and the displacement of all aircraft; and constructing a target loss function based on the stage loss function, the value function loss function, and the policy entropy loss function.

[0025] For example, the Clipped Surrogate Objective can be expressed as: ; in, is the truncation loss function, are the parameters of the policy network (such as the weights of the neural network), n is the batch index (indicating the nth policy update), represents the expected time step, represents the action probability ratio at time t, For the current strategy in state Next select action The probability of The old policy before the update is in the state Select Action The probability of is the advantage function, is the shear function, is the shear coefficient.

[0026] The shear function and shear coefficient are used to limit the difference between the new and old strategies, and are functions used to describe the advantages relative to other actions.

[0027] The value function loss function can be expressed as: ; in, is the value function loss function, Indicates that the status The state value under (that is, the output vector), , is the reward at time t, represents the value function at time t, , is the number of moves of all aircraft, is the average maintenance bay utilization rate, is the displacement of all aircraft, when the discount factor When the total long-term return , represents the final goal after all maintenance tasks are completed. Learning maximization The strategy is equivalent to minimizing the value function loss function; ; ; ; Indicates aircraft Maintenance items Whether it is assigned to a maintenance stand (If assigned, ,otherwise ). Indicates the aircraft Related maintenance items The maximum completion time, where It's with the plane The total number of associated maintenance items, m represents the total number of maintenance bays. Indicates aircraft From current location Move to the target location distance.

[0028] The policy entropy loss function can be expressed as: ; in, represents the policy entropy loss function, Status action space.

[0029] The target loss can be expressed as: ; in, is the target loss function, N is the total number of batches included in a training cycle, is the first hyperparameter, is the second hyperparameter, is the third hyperparameter, 、 as well as Both can be set based on experience.

[0030] In one embodiment, it is possible to set , as well as Setting hyperparameters allows the agent to maintain a stable learning process, balance long-term and short-term benefits, and effectively balance exploration and exploitation. Hyperparameters can be pre-adjusted on a small-scale instance based on experience, and then fixed on instances of all other sizes.

[0031] Step S140 , based on the original feature vector and the first heterogeneous graph model, a maintenance bay scheduling prediction network is trained with the goal of minimizing the target loss function to obtain a target scheduling network.

[0032] In one embodiment of the present application, based on the original feature vector and the first heterogeneous graph model, the maintenance aircraft scheduling prediction network is trained with the goal of minimizing the target loss function to obtain the target scheduling network, which may include steps S141 to S145.

[0033] In step S141, the maintenance project feature vector and the first heterogeneous graph model are taken as input, and a maintenance project node embedding vector is obtained based on the maintenance project graph attention network.

[0034] In one embodiment of the present application, the process of taking the maintenance project feature vector and the first heterogeneous graph model as input and obtaining the maintenance project node embedding vector based on the maintenance project graph attention network may include steps S1411 to S1413.

[0035] Step S1411: determining a first attention coefficient between each maintenance item and its upstream maintenance item and downstream maintenance item based on the maintenance item feature vector and the first heterogeneous graph model.

[0036] In one embodiment of the present application, the attention coefficient between each maintenance project and its upstream maintenance project and downstream maintenance project can be calculated based on the input original feature vector and the first heterogeneous graph model, that is, the set of all maintenance project nodes that have connection arcs with the maintenance project node includes the maintenance project node itself.

[0037] To embed maintenance project nodes, the maintenance project feature vectors are used as input to the maintenance project graph attention network. The input dimension of the subsequent core function layer is the same as the output dimension of the previous core function layer. The attention mechanism aggregates information from surrounding neighboring nodes to extract features for each maintenance project node.

[0038] In one embodiment, to better capture deep-level features, the maintenance project node embedding vectors are calculated using a multi-layer stacking approach. To implement this multi-layer stacking approach, the maintenance project graph attention network includes multiple layers of graph attention networks. Except for the first layer, for each layer of graph attention network, the output of the previous layer is used as the input of the next layer.

[0039] In one embodiment, step S1411 can be expressed as: Maintenance project nodes The embedding process aims to identify the most important maintenance items through their inherent properties, thereby The embedding process links different maintenance items in the same aircraft to be inspected. Each maintenance project node in the input feature , first calculate each maintenance project node and its upstream maintenance projects and downstream maintenance projects The first attention coefficient between: ; in, is the first attention coefficient, is the first linear transformation matrix, for all The maintenance items are linearly transformed. is the second linear transformation matrix; LeakyReLu represents the activation function; Represents the input maintenance item node embedding vector.

[0040] For example, please refer to Figure 2 , for maintenance project nodes , its upstream maintenance project nodes and downstream maintenance project nodes are and .

[0041] Step S1412: normalize each first attention coefficient to obtain multiple second attention coefficients.

[0042] During the calculation process, since the upstream maintenance items or downstream maintenance items of some maintenance items may not exist or will be cancelled at a certain time step, the attention coefficients of the above maintenance items are dynamically masked. Apply the SoftMax function to obtain a set of normalized second attention coefficients .

[0043] Step S1413: Obtain a maintenance item node embedding vector based on all second attention coefficients and all maintenance item feature vectors.

[0044] Combine the second attention coefficient with the transformed input features 、 and The maintenance project node embedding vector is obtained after linear weighted combination .

[0045] In one embodiment, step 303 may be expressed as: ; in, Represents the first-level maintenance project node After the iteration of multiple layers of networks, based on the graph attention mechanism, the output of the last layer of graph attention network is the embedding vector of the maintenance project node.

[0046] By sequentially connecting multiple maintenance project core function layers, the maintenance project nodes can be The information is transmitted to the aircraft to be repaired In the above process, there is no need to calculate the embedding of the two virtual maintenance project nodes Start and End for all maintenance projects. Through the embodiment of step S141, the priority constraint relationship of the maintenance projects is focused on, so that the embedded project node characteristics can reflect the temporal sequence and correlation between projects, further improving the rationality of the scheduling strategy.

[0047] In step S142, the original feature vector and the first heterogeneous graph model are used as input, and an embedding vector of the maintenance bay node is obtained based on the maintenance bay graph attention network.

[0048] In one embodiment of the present application, the process of taking the original feature vector and the first heterogeneous graph model as input and obtaining the maintenance bay node embedding vector based on the maintenance bay graph attention network may include: Step S1421 : determining a maintenance stand node feature based on the maintenance stand feature vector and the maintenance stand-maintenance stand arc feature vector in the original feature vector.

[0049] In one embodiment of the present application, the process of determining the maintenance stand node feature based on the maintenance stand feature vector and the maintenance stand-maintenance stand arc feature vector in the original feature vector may include: splicing the maintenance stand feature vector and the maintenance stand-maintenance stand arc feature vector to obtain the maintenance stand node feature.

[0050] For example, each maintenance bay By taking its original features With the corresponding Original characteristics of the arc (maintenance stand-maintenance stand arc) Connect and expand to form the maintenance station node feature after gluing .

[0051] It should be noted that the maintenance stand-maintenance stand arc may be the distance between adjacent wingtips of two aircraft.

[0052] To embed maintenance bay nodes, the original feature vector designed previously is used as the input to the maintenance bay graph attention network. The input dimension of the subsequent core function layer is the same as the output dimension of the previous core function layer. The attention mechanism aggregates information from surrounding neighboring nodes to extract features for each maintenance bay node.

[0053] In one embodiment, to better capture deep-level features, the maintenance bay node embedding vector is calculated using a multi-layer stacking approach. To implement this multi-layer stacking approach, in one embodiment, the maintenance bay graph attention network includes multiple layers of graph attention networks. Except for the first layer, for each layer of the graph attention network, the output of the previous layer is used as the input of the next layer.

[0054] In step S1422, based on the original feature vector, the maintenance station node feature and the first heterogeneous graph model, the self-attention coefficient of any maintenance station is obtained, and the third attention coefficient between all maintenance items that have a maintenance item-maintenance station arc with any maintenance station is obtained.

[0055] Based on the original feature vector, maintenance station node features and the first heterogeneous graph model, maintenance projects are effectively , maintenance bay , Maintenance Project-Maintenance Stand Arc and the maintenance bay-maintenance bay arc The information is integrated into the calculation process of the attention mechanism.

[0056] In one embodiment, step S1422 can be expressed as: ; ; in, and Represents different linear transformation matrices, is the self-attention coefficient, The third attention coefficient is as follows: In the process of determining the aircraft maintenance stand scheduling strategy, adjacent maintenance stands must follow spatial constraints to ensure a safe distance. , and are different linear transformation matrices. In this way, the maintenance items are effectively , maintenance bay And maintenance project - maintenance stand arc The information is integrated into the calculation process of the attention mechanism.

[0057] In step S1423, the self-attention coefficient and the third attention coefficient are normalized respectively to obtain a first normalization coefficient and a second normalization coefficient.

[0058] Use the SoftMax function to transform all and Normalization gets the first normalization coefficient and the second normalization coefficient .

[0059] Step S1424: Obtain a maintenance station node embedding vector based on the first normalization coefficient, the second normalization coefficient, the maintenance station node feature, and the maintenance item-maintenance station arc feature vector.

[0060] The activation function ELU is used on the transformed input features to obtain the maintenance station node embedding vector : ; in, Indicates the first-level maintenance bay node The embedding vector of Indicates maintenance bay Neighborhood Repair Project After the iteration of multiple layers of networks, based on the graph attention mechanism, the output of the last layer of the maintenance station graph attention network is the maintenance station node embedding vector.

[0061] In one embodiment, to better capture deep-level features, the maintenance bay node embedding vector is calculated using a multi-layer stacking approach. To implement this multi-layer stacking approach, in one embodiment, the maintenance bay graph attention network includes multiple layers of graph attention networks. Except for the first layer, for each layer of the graph attention network, the output of the previous layer is used as the input of the next layer.

[0062] It should be noted that both the maintenance project map attention network and the maintenance aircraft position map attention network include multi-layer attention networks.

[0063] Step S143: concatenate the maintenance project node embedding vector and the maintenance position node embedding vector to obtain a global feature vector.

[0064] After being processed by the L-layer maintenance aircraft flexible scheduling graph attention network core layer, the output results are average pooled and the learned features and It is used in subsequent decision-making actions to connect the results to form the global feature vector of the maintenance bay scheduling instance. .

[0065] In one embodiment, step S143 can be expressed as: ; in, represents the global eigenvector, Represents the maintenance project node output by the attention network of the maintenance project graph The embedding vector of A maintenance bay node representing the output of the maintenance bay graph attention network The embedding vector of Represents a collection of maintenance project nodes, Indicates the number of maintenance bay nodes. Represents a vector concatenation operation.

[0066] This step achieves further feature extraction and converts heterogeneous graph models of varying sizes into fixed-dimensional and Obtaining state embedding helps the network better understand the complex relationships between maintenance bays and maintenance items.

[0067] Step S144, concatenate the maintenance project node embedding vector, the maintenance station node embedding vector, the global feature vector, the maintenance project-maintenance station arc feature vector, and the maintenance station-maintenance station arc feature vector to obtain a concatenated vector, and input the concatenated vector into the first multi-layer perceptron to obtain multiple action scalars.

[0068] In one embodiment, step S144 can be expressed as: ; Specifically, for each feasible action , and the related maintenance bay node characteristics , Maintenance project node characteristics , global state characteristics , Maintenance Project-Maintenance Position Arc Characteristics and the maintenance stand-maintenance stand arc feature vector are glued and fed into the MLP to obtain its Selected action scalar .in There are two hidden layers.

[0069] Step S145 , normalizing all action scalars to obtain an aircraft maintenance stand scheduling strategy.

[0070] For all The SoftMax normalization function is used to calculate the probability distribution of actions.

[0071] In one embodiment, step S145 can be expressed as: ; Step S146: input the global feature vector into the second multi-layer perceptron to obtain an output vector output by the second multi-layer perceptron.

[0072] The target loss function is back-propagated to the maintenance project graph attention network and the maintenance aircraft stand graph attention network to realize the training of the maintenance aircraft stand scheduling prediction network.

[0073] A random scheduling strategy is generated based on the first multilayer perceptron, and the results of the second multilayer perceptron are used to evaluate the generated scheduling strategy. By using the two multilayer perceptrons as the action decision layer and the action evaluation layer, respectively, a reinforcement learning algorithm for the action decision layer of the actor-critic framework is constructed, thereby achieving training for the maintenance bay scheduling prediction network.

[0074] For example, see Figure 3 , which is an architectural diagram of a maintenance bay scheduling prediction network, illustrated in an exemplary embodiment of this application. An OM arc represents a maintenance item-to-maintenance bay arc, and an MM arc represents a maintenance bay-to-maintenance bay arc. Leveraging the characteristics and advantages of heterogeneous graph models, a graph neural network framework specifically designed for the maintenance bay scheduling problem is introduced. This framework cleverly captures the characteristic representations of nodes and arcs.

[0075] Step S150 , obtaining input feature vectors, the input feature vectors including feature vectors of multiple maintenance items of the aircraft to be maintained, feature vectors of multiple maintenance bays, feature vectors of maintenance item-maintenance bay arcs, and feature vectors of maintenance bay-maintenance bay arcs.

[0076] Step S160: construct a second heterogeneous graph model based on the input feature vector.

[0077] Step S170, inputting the input feature vector and the second heterogeneous graph model into the target scheduling network to obtain an aircraft maintenance stand scheduling strategy determination strategy, so as to schedule the aircraft to be maintained based on the aircraft maintenance stand scheduling strategy determination strategy; the target scheduling network is used to obtain an output aircraft maintenance stand scheduling strategy determination strategy based on the input feature vector and the second heterogeneous graph model.

[0078] Figure 4 FIG. 1 is a block diagram of an exemplary embodiment of the present invention showing an apparatus for determining an aircraft maintenance stand scheduling strategy. Figure 4 As shown, the exemplary aircraft maintenance stand scheduling strategy determination device 400 includes: The first data acquisition module 410 is configured to acquire original feature vectors, wherein the original feature vectors include multiple maintenance item feature vectors, multiple maintenance bay feature vectors, a maintenance item-maintenance bay arc feature vector, and a maintenance bay-maintenance bay arc feature vector. The maintenance item-maintenance bay arc feature vector includes the maintenance time for each maintenance item at each available maintenance bay. The maintenance bay-maintenance bay arc feature vector includes the distance between each maintenance bay and its adjacent maintenance bay. A first processing module 420 is configured to construct a first heterogeneous graph model based on the original feature vector; the first heterogeneous graph model includes maintenance project nodes, maintenance station nodes, connection arcs between each maintenance project and maintenance station, and connection arcs between each maintenance station; A function construction module 430 is used to construct a target loss function based on the number of aircraft movements, the average maintenance bay utilization rate, and the displacement of all aircraft; A model training module 440 is configured to train a maintenance bay scheduling prediction network based on the original feature vector and the first heterogeneous graph model with the goal of minimizing a target loss function to obtain a target scheduling network; A second data acquisition module 450 is configured to acquire input feature vectors, the input feature vectors including a plurality of maintenance item feature vectors of the aircraft to be maintained, a plurality of maintenance stand feature vectors, a feature vector of a maintenance item-maintenance stand arc, and a feature vector of a maintenance stand-maintenance stand arc; A second processing module 460 is configured to construct a second heterogeneous graph model based on the input feature vector; A strategy acquisition module 470 is configured to input the input feature vector and the second heterogeneous graph model into the target scheduling network to obtain an aircraft maintenance stand scheduling strategy determination strategy, so as to schedule aircraft to be maintained based on the aircraft maintenance stand scheduling strategy determination strategy; the target scheduling network is configured to obtain an output aircraft maintenance stand scheduling strategy determination strategy based on the input feature vector and the second heterogeneous graph model.

[0079] In one embodiment of the present application, the function building block includes: The first function construction unit is used to construct a value function loss function based on the number of aircraft movements, the average maintenance bay utilization rate, and the displacement of all aircraft; The second function construction unit is used to construct a target loss function based on the stage loss function, the value function loss function and the strategy entropy loss function.

[0080] In one embodiment of the present application, the model training module includes: a first data processing unit, configured to take the maintenance item feature vector and the first heterogeneous graph model as input, and obtain a maintenance item node embedding vector based on a maintenance item graph attention network; A second data processing unit is configured to take the original feature vector and the first heterogeneous graph model as input, and obtain a maintenance bay node embedding vector based on a maintenance bay graph attention network; A third data processing unit is used to concatenate the maintenance project node embedding vector and the maintenance position node embedding vector to obtain a global feature vector; a fourth data processing unit, configured to concatenate the maintenance project node embedding vector, the maintenance position node embedding vector, the global feature vector, the maintenance project-maintenance position arc feature vector, and the maintenance position-maintenance position arc feature vector to obtain a concatenated vector, and input the concatenated vector into a multilayer perceptron to obtain a plurality of action scalars; The fifth data processing unit is used to normalize all action scalars to obtain the probability distribution of the action and obtain the aircraft maintenance stand scheduling strategy.

[0081] It should be noted that the aircraft maintenance stand scheduling strategy determination apparatus provided in the above-mentioned embodiment and the aircraft maintenance stand scheduling strategy determination method provided in the above-mentioned embodiment are based on the same concept. The specific manner in which each module and unit performs operations has been described in detail in the method embodiment and will not be repeated here. In actual applications, the aircraft maintenance stand scheduling strategy determination apparatus provided in the above-mentioned embodiment can, as needed, allocate the above-mentioned functions to different functional modules, i.e., divide the internal structure of the system into different functional modules to perform all or part of the functions described above, and this is not a limitation herein.

[0082] An embodiment of the present application also provides an electronic device, comprising: one or more processors; a storage device for storing one or more programs, wherein when the one or more programs are executed by the one or more processors, the electronic device implements the aircraft maintenance stand scheduling strategy determination method provided in each of the above embodiments.

[0083] Another aspect of the present application provides a computer-readable storage medium storing a computer program. When executed by a computer processor, the computer program causes the computer to perform the aircraft maintenance stand scheduling strategy determination method provided in each of the aforementioned embodiments. The computer-readable storage medium may be included in the electronic device described in the aforementioned embodiments, or may exist independently and not be incorporated into the electronic device.

[0084] Another aspect of the present application provides a computer program product or computer program, comprising computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform the aircraft maintenance stand scheduling strategy determination method provided in each of the above-described embodiments.

[0085] In the embodiments of this application, the terms "first" and "second" are used for descriptive purposes only and should not be understood to indicate or imply relative importance. Throughout the specification and claims, the terms "including" and "comprising" are open-ended terms and should be interpreted as "including but not limited to."

[0086] The above embodiments are merely illustrative of the principles and effects of this application and are not intended to limit this application. Anyone skilled in the art may modify or alter the above embodiments without departing from the spirit and scope of this application. Therefore, any equivalent modifications or alterations accomplished by a person of ordinary skill in the art without departing from the spirit and technical concepts disclosed in this application shall be covered by the claims of this application.

Claims

1. A method for determining an aircraft maintenance stand scheduling strategy, characterized in that: include: Obtaining original feature vectors, the original feature vectors including multiple maintenance item feature vectors, multiple maintenance bay feature vectors, a maintenance item-maintenance bay arc feature vector, and a maintenance bay-maintenance bay arc feature vector; the maintenance item-maintenance bay arc feature vector including the maintenance time for each maintenance item at each available maintenance bay; and the maintenance bay-maintenance bay arc feature vector including the distance between each maintenance bay and its adjacent maintenance bay; Constructing a first heterogeneous graph model based on the original feature vector; the first heterogeneous graph model includes maintenance project nodes, maintenance position nodes, connection arcs between each maintenance project and maintenance position, and connection arcs between each maintenance position; Construct a target loss function based on the number of aircraft moves, average maintenance bay utilization, and the displacement of all aircraft; Based on the original feature vector and the first heterogeneous graph model, training a maintenance aircraft bay scheduling prediction network with the goal of minimizing a target loss function to obtain a target scheduling network; Obtaining an input feature vector, the input feature vector including a plurality of maintenance item feature vectors of the aircraft to be maintained, a plurality of maintenance stand feature vectors, a feature vector of a maintenance item-maintenance stand arc, and a feature vector of a maintenance stand-maintenance stand arc; constructing a second heterogeneous graph model based on the input feature vector; The input feature vector and the second heterogeneous graph model are input into the target scheduling network to obtain an aircraft maintenance stand scheduling strategy determination strategy, so as to schedule the aircraft to be maintained based on the aircraft maintenance stand scheduling strategy determination strategy; the target scheduling network is used to obtain an output aircraft maintenance stand scheduling strategy based on the input feature vector and the second heterogeneous graph model.

2. The method for determining aircraft maintenance stand scheduling strategy according to claim 1, wherein: Based on the number of aircraft movements, the average maintenance bay utilization rate, and the displacement of all aircraft, a target loss function is constructed, including: Construct a value function loss function based on the number of aircraft movements, average maintenance bay utilization, and the displacement of all aircraft; Based on the stage loss function, value function loss function and policy entropy loss function, the target loss function is constructed.

3. The method for determining aircraft maintenance stand scheduling strategy according to claim 1, wherein: Based on the original feature vector and the first heterogeneous graph model, a maintenance aircraft bay scheduling prediction network is trained with the goal of minimizing the target loss function to obtain a target scheduling network, including: Taking the maintenance project feature vector and the first heterogeneous graph model as input, obtaining a maintenance project node embedding vector based on a maintenance project graph attention network; Taking the original feature vector and the first heterogeneous graph model as input, obtaining a maintenance bay node embedding vector based on a maintenance bay graph attention network; The global feature vector is obtained by concatenating the embedding vector of the maintenance project node and the embedding vector of the maintenance position node; splicing the maintenance project node embedding vector, the maintenance station node embedding vector, the global feature vector, the maintenance project-maintenance station arc feature vector, and the maintenance station-maintenance station arc feature vector to obtain a splicing vector, and inputting the splicing vector into the first multi-layer perceptron to obtain multiple action scalars; Normalize all action scalars to obtain the aircraft maintenance bay scheduling strategy; The global feature vector is input into the second multi-layer perceptron to obtain an output vector output by the second multi-layer perceptron.

4. The method for determining aircraft maintenance stand scheduling strategy according to claim 3, wherein: Taking the maintenance item feature vector and the first heterogeneous graph model as input, obtaining a maintenance item node embedding vector based on the maintenance item graph attention network includes: Determining a first attention coefficient between each maintenance item and its upstream maintenance item and downstream maintenance item based on the maintenance item feature vector and the first heterogeneous graph model; Normalize each first attention coefficient to obtain multiple second attention coefficients; Based on all the second attention coefficients and all the maintenance item feature vectors, the maintenance item node embedding vector is obtained.

5. The method for determining aircraft maintenance stand scheduling strategy according to claim 3, wherein: Taking the original feature vector and the first heterogeneous graph model as input, and based on the maintenance bay graph attention network, obtaining the maintenance bay node embedding vector includes: Determine the maintenance stand node feature based on the maintenance stand feature vector and the maintenance stand-maintenance stand arc feature vector in the original feature vector; Based on the original feature vector, the maintenance station node feature, and the first heterogeneous graph model, a self-attention coefficient of any maintenance station is obtained, and a third attention coefficient between all maintenance items that have a maintenance item-maintenance station arc with the maintenance station is obtained; Normalize the self-attention coefficient and the third attention coefficient respectively to obtain the first normalized coefficient and the second normalized coefficient; Based on the first normalization coefficient, the second normalization coefficient, the maintenance stand node features, and the maintenance item-maintenance stand arc feature vector, the maintenance stand node embedding vector is obtained.

6. The method for determining aircraft maintenance stand scheduling strategy according to claim 5, characterized in that: Determining a maintenance stand node feature based on the maintenance stand feature vector and the maintenance stand-maintenance stand arc feature vector in the original feature vector includes: The maintenance stand feature vector and the maintenance stand-maintenance stand arc feature vector are concatenated to obtain the maintenance stand node feature.

7. The method for determining aircraft maintenance stand scheduling strategy according to claim 3, wherein: The maintenance project map attention network and the maintenance machine location map attention network both include multi-layer attention networks.

8. A device for determining aircraft maintenance stand scheduling strategy, characterized in that: include: A first data acquisition module is configured to acquire original feature vectors, wherein the original feature vectors include a plurality of maintenance item feature vectors, a plurality of maintenance bay feature vectors, a maintenance item-maintenance bay arc feature vector, and a maintenance bay-maintenance bay arc feature vector. The maintenance item-maintenance bay arc feature vector includes the maintenance time for each maintenance item at each available maintenance bay. The maintenance bay-maintenance bay arc feature vector includes the distance between each maintenance bay and its adjacent maintenance bay. A first processing module is configured to construct a first heterogeneous graph model based on the original feature vector; the first heterogeneous graph model includes maintenance project nodes, maintenance station nodes, connection arcs between each maintenance project and maintenance station, and connection arcs between each maintenance station; A function building module is used to build a target loss function based on the number of aircraft movements, the average maintenance bay utilization rate, and the displacement of all aircraft; A model training module, configured to train a maintenance bay scheduling prediction network based on the original feature vector and the first heterogeneous graph model with the goal of minimizing a target loss function to obtain a target scheduling network; a second data acquisition module, configured to acquire an input feature vector, wherein the input feature vector includes a plurality of maintenance item feature vectors of the aircraft to be maintained, a plurality of maintenance stand feature vectors, a feature vector of a maintenance item-maintenance stand arc, and a feature vector of a maintenance stand-maintenance stand arc; A second processing module, configured to construct a second heterogeneous graph model based on the input feature vector; A strategy acquisition module is configured to input the input feature vector and the second heterogeneous graph model into the target scheduling network to obtain an aircraft maintenance stand scheduling strategy determination strategy, so as to schedule the aircraft to be maintained based on the aircraft maintenance stand scheduling strategy determination strategy; the target scheduling network is configured to obtain an output aircraft maintenance stand scheduling strategy determination strategy based on the input feature vector and the second heterogeneous graph model.

9. The aircraft maintenance stand scheduling strategy determination device according to claim 8, characterized in that: Function building blocks include: The first function construction unit is used to construct a value function loss function based on the number of aircraft movements, the average maintenance bay utilization rate, and the displacement of all aircraft; The second function construction unit is used to construct a target loss function based on the stage loss function, the value function loss function and the strategy entropy loss function.

10. The aircraft maintenance stand scheduling strategy determination device according to claim 8, characterized in that: The model training module includes: a first data processing unit, configured to take the maintenance item feature vector and the first heterogeneous graph model as input, and obtain a maintenance item node embedding vector based on a maintenance item graph attention network; A second data processing unit is configured to take the original feature vector and the first heterogeneous graph model as input, and obtain a maintenance bay node embedding vector based on a maintenance bay graph attention network; A third data processing unit is used to concatenate the maintenance project node embedding vector and the maintenance position node embedding vector to obtain a global feature vector; a fourth data processing unit, configured to concatenate the maintenance project node embedding vector, the maintenance position node embedding vector, the global feature vector, the maintenance project-maintenance position arc feature vector, and the maintenance position-maintenance position arc feature vector to obtain a concatenated vector, and input the concatenated vector into a multilayer perceptron to obtain a plurality of action scalars; The fifth data processing unit is used to normalize all action scalars to obtain the probability distribution of the action and obtain the aircraft maintenance stand scheduling strategy.

Images