Time length prediction method of target time period, equipment and medium

Abstract

The invention provides a duration prediction method and device for a target time period and a medium, and relates to the technical field of data processing, and the method comprises the steps: building an initial duration prediction model based on a preset task feature tag list, a progressive hierarchical extraction framework and a mixed Bayesian neural network; training the initial time length prediction model based on a preset training data set, optimizing the initial time length prediction model based on a joint loss optimization method to obtain a target time length prediction model, and inputting a feature value list corresponding to the flight task of the to-be-predicted aircraft into the target time length prediction model to obtain a target prediction time length list, it can be known that the target time length prediction model provided by the invention combines the progressive hierarchical extraction framework and the mixed Bayesian neural network, has strong feature extraction capability, can simultaneously predict the time lengths of multiple target time periods, and is beneficial for improving the efficiency and accuracy of time length prediction.

Description

Technical Field

[0001] The present invention relates to the field of data processing technology, and in particular to a method, device and medium for predicting the duration of a target period. Background Art

[0002] The preparatory actions that need to be performed before the execution of an aircraft flight mission are crucial. The time period (target period) between adjacent preparatory actions can directly affect the overall efficiency and safety of the aircraft flight mission. Therefore, accurately predicting the duration of the target period is of great significance for optimizing the evaluation indicators of the aircraft flight mission. The duration prediction methods in the prior art are mainly rule-based methods and data-driven methods. Among them, the rule-based method predicts the duration and relies on fixed rules summarized from manual experience and historical data. The data-driven method predicts the duration by using machine learning and statistical models to learn rules from historical data for prediction.

[0003] However, the above method also has the following technical problems:

[0004] The rules summarized based on manual experience and historical data are greatly affected by human factors, are too simple and unstable. When predicting duration based on data-driven methods, commonly used models include linear regression, decision trees, random forests and support vector machines. However, commonly used models are relatively simple in feature extraction. When faced with new and unprecedented data, the prediction performance of the model will decrease, and existing models are difficult to take into account multiple prediction targets at the same time. They can only predict the duration of one target period at the same time. Therefore, the efficiency of duration prediction based on the above method is low, and the accuracy of the predicted duration is low. Summary of the invention

[0005] In view of the above technical problems, the technical solution adopted by the present invention is:

[0006] According to a first aspect of the present invention, a method for predicting the duration of a target period is provided, wherein the target period is a time period between adjacent preset execution actions, and the preset execution action is a preparatory action that needs to be performed before an aircraft flight mission is performed. The method comprises the following steps:

[0007] S1. Establish an initial duration prediction model based on the preset task feature label list F, the progressive hierarchical extraction framework and the hybrid Bayesian neural network, where F = {F1, F2, ..., F i , ..., F m}, F i is the preset task feature label of the i-th preset task feature of the aircraft flight mission, the value of i ranges from 1 to m, m is the number of preset task feature labels, and the input of the initial duration prediction model is the feature value list A corresponding to the aircraft flight mission = {A1, A2, ..., Ai , ..., A m}, A i For A and F i The output of the initial duration prediction model is the predicted duration list B = {B1, B2, ..., B j , ..., B n}, B j is the predicted duration obtained after the j-th prediction task predicts the j-th target time period, j ranges from 1 to n, n is the number of prediction tasks, the initial duration prediction model includes a first feature extraction module and a second feature extraction module, the first feature extraction module is used to extract shallow features, the second feature extraction module is used to extract deep features, the structures of the first feature extraction module and the second feature extraction module are both progressive hierarchical extraction structures, the first feature extraction module and the second feature extraction module are both composed of n first-class networks and several second-class networks, the first-class network is a neural network used to extract specific features of a single prediction task, the second-class network is a neural network used to extract common features of all prediction tasks, and the first-class network and the second-class network are both built based on the Bayesian neural network structure.

[0008] S2. Train the initial duration prediction model based on a preset training data set and optimize the initial duration prediction model based on a joint loss optimization method to obtain a target duration prediction model.

[0009] S3, input the feature value list TZ corresponding to the flight mission to be predicted into the target duration prediction model to obtain the target prediction duration list SC, where TZ = {TZ1, TZ2, ..., TZ i , ..., TZ m}, TZ i For TZ and F i The corresponding eigenvalues, SC = {SC1, SC2, ..., SC j , ..., SC n}, SC j is the predicted duration of the jth target period of the flight mission of the aircraft to be predicted.

[0010] According to a second aspect of the present invention, a non-transitory computer-readable storage medium is provided, in which a computer program is stored. The computer program is loaded and executed by a processor to implement the aforementioned method.

[0011] According to a third aspect of the present invention, there is provided an electronic device, comprising: a processor, a memory, and a computer program stored in the memory and executable on the processor, wherein the processor implements the aforementioned method when executing the computer program.

[0012] The present invention has at least the following beneficial effects:

[0013] The present invention provides a method, device and medium for predicting the duration of a target time period. The method establishes an initial duration prediction model based on a preset task feature label list, a progressive hierarchical extraction framework and a hybrid Bayesian neural network, trains the initial duration prediction model based on a preset training data set and optimizes the initial duration prediction model based on a joint loss optimization method to obtain a target duration prediction model, inputs a feature value list corresponding to a flight mission of an aircraft to be predicted into the target duration prediction model to obtain a target prediction duration list. It can be seen that the target duration prediction model proposed by the present invention combines the progressive hierarchical extraction framework and the hybrid Bayesian neural network, has a strong feature extraction capability and can simultaneously predict the durations of multiple target time periods, which is beneficial to improving the efficiency and accuracy of duration prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0015] Figure 1 A flowchart of a method for predicting the duration of a target time period provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0016] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative work are within the scope of protection of the present invention.

[0017] It should be noted that the terms "first", "second", etc. in the specification and claims of the present invention and the above-mentioned drawings are used to distinguish similar tasks, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged where appropriate, so that the embodiments of the present invention described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, system, product, or server that includes a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products, or devices.

[0018] An embodiment of the present invention provides a method for predicting the duration of a target period, wherein the target period is a time period between adjacent preset execution actions, and the preset execution action is a preparatory action that needs to be performed before the aircraft flight mission is executed. The method comprises the following steps: Figure 1 As shown:

[0019] S1. Establish an initial duration prediction model based on the preset task feature label list F, the progressive hierarchical extraction framework and the hybrid Bayesian neural network, where F = {F1, F2, ..., F i , ..., F m}, F i is the preset task feature label of the i-th preset task feature of the aircraft flight mission, the value of i ranges from 1 to m, m is the number of preset task feature labels, and the input of the initial duration prediction model is the feature value list A corresponding to the aircraft flight mission = {A1, A2, ..., A i , ..., A m}, A i For A and F i The output of the initial duration prediction model is the predicted duration list B = {B1, B2, ..., B j , ..., B n}, B j The predicted duration is obtained after the j-th prediction task predicts the j-th target period. The value of j ranges from 1 to n, and n is the number of prediction tasks.

[0020] Specifically, the prediction tasks correspond to the target time periods one by one.

[0021] In a specific embodiment, the aircraft mission is a flight.

[0022] Specifically, the preset mission characteristics of an aircraft flight mission are characteristics that need to be set by technical personnel in this field according to actual needs. For example, when the aircraft flight mission is a flight, the preset mission characteristics of the aircraft flight mission are aircraft model, aircraft stand number, aircraft stand type, route nature, arrival deviation, arrival flight density, arrival time period, departure time period, minimum transit time, number of departing passengers, number of departing luggage pieces, departing luggage weight, etc.

[0023] Specifically, the initial duration prediction model includes a first feature extraction module and a second feature extraction module. The first feature extraction module is used to extract shallow features, and the second feature extraction module is used to extract deep features. Shallow features can be understood as features extracted from the shallower layers of the network, and deep features can be understood as features extracted from the deeper layers of the network.

[0024] Specifically, the structures of the first feature extraction module and the second feature extraction module are both progressive hierarchical extraction structures.

[0025] Specifically, the first feature extraction module and the second feature extraction module are both composed of n first-type networks and a plurality of second-type networks.

[0026] Specifically, the first type of network is a neural network used to extract specific features of a single prediction task.

[0027] Specifically, the second type of network is a neural network used to extract common features for all prediction tasks.

[0028] Specifically, the first type of network corresponds one-to-one to the prediction task.

[0029] Specifically, both the first type of network and the second type of network are constructed based on the Bayesian neural network structure, which can be further understood as: the first feature extraction module and the second feature extraction module are both integrated with the Bayesian neural network.

[0030] Specifically, the initial duration prediction model also includes n task towers.

[0031] Specifically, the task towers correspond one to one with the prediction tasks.

[0032] Specifically, the task tower is built based on a fully connected network structure.

[0033] Specifically, the first feature extraction module and the second feature extraction module are in communication connection with each task tower.

[0034] Specifically, in the initial duration prediction model, the self-attention layer is used to obtain the attention weight list set C corresponding to A = {C1, C2, ..., C i , ..., C m} and obtain the intermediate feature vector list E corresponding to A based on C = {E1, E2, ..., E i ,……,E m}, where C i A i The corresponding attention weight list, E i A i The corresponding intermediate feature vector includes the following steps:

[0035] Get A i The corresponding query vector Q i , key vector K i Sum value vector V i , Q i , K i and V i Meet the following conditions respectively:

[0036] Q i =W Q ×A i , K i=W K ×A i , V i =W V ×A i , W Q is the weight matrix used to calculate the query vector, W K is the weight matrix used to calculate the key vector, W V is the weight matrix corresponding to the value vector used to calculate.

[0037] Get C i ={C i1 , C i2 , ..., C iα , ..., C im}, C iα Q i K α The attention weight, K α A α The corresponding key vector, A α For A and F α The corresponding eigenvalue, F α is the preset task feature label of the αth preset task feature of the aircraft flight mission, where α ranges from 1 to m, and C iα Meet the following conditions:

[0038] C iα =softmax(Q i ×K α T / d k 1 / 2 ), where K α T K α The transposed matrix, d k is the dimension of the key vector.

[0039] Get E i , where E i Meet the following conditions:

[0040] E i =∑ m α＝1 (C iα ×V α ), V α A α The corresponding value vector.

[0041] Through the above steps, the attention of the initial duration prediction model to the characteristic values ​​corresponding to the aircraft flight mission is dynamically adjusted based on the self-attention mechanism to obtain a set of attention weight lists corresponding to the characteristic value list corresponding to the aircraft flight mission, and the intermediate feature vector list corresponding to the characteristic value list corresponding to the aircraft flight mission is obtained according to the attention weight list set. This can capture the relationship between the characteristic values ​​corresponding to the aircraft flight mission and the target time period, which is conducive to improving the interpretability of the initial duration prediction model.

[0042] Specifically, during the operation of the initial duration prediction model, B is obtained through the following steps: j :

[0043] Get the j-th prediction task G according to E j The corresponding selection matrix H j , wherein those skilled in the art know that any method for selecting a matrix based on a feature vector acquisition task in the prior art falls within the protection scope of the present invention and will not be described in detail herein.

[0044] According to A i and H j Get G j Middle A i The corresponding first feature weight R ij , R ij Meet the following conditions:

[0045] R ij =A i ×H j , the first feature weight is used to indicate the importance of the feature value for the prediction task.

[0046] According to R ij Get G j The corresponding comprehensive feature weight L j , L j Meet the following conditions:

[0047] L j =∑ m i＝1 (exp(M(A i )) / ∑ m i＝1 exp(M(A i ))×R ij ), where M(A i ) is the key module for A i The second feature weight obtained after processing, the key module is the module containing the gated recurrent unit (GRU) and the attention mechanism in the initial duration prediction model, and the key module is used to dynamically select or weight specific time steps or features from the input sequence.

[0048] L j Input to G j Get B in the corresponding mission tower j .

[0049] Through the above steps, during the operation of the initial duration prediction model, the predicted duration is determined based on the attention gating network, which is conducive to improving the accuracy of the obtained predicted duration.

[0050] Specifically, the final predicted distribution FB of the task tower on the sample data of group S meets the following conditions:

[0051] FB = p(B|A, D), p() is the likelihood function;

[0052] p(B|A,D)≈1 / S×∑ S f＝1 p(B|A,ω f );

[0053] ω f =μ f +λ f ×(log(1+exp(ρ f ))) 1 / 2 , where D is a historical data set consisting of S groups of sample data, (μ f , ρ f ) is the variational parameter corresponding to the f-th group of sample data, the value of f ranges from 1 to S, λ f is the weight parameter corresponding to the fth group of sample data. Those skilled in the art know that the sample data are data predetermined by those skilled in the art according to actual needs and will not be described in detail here.

[0054] Specifically, f It follows a normal distribution with a mean of 0 and a variance of 1.

[0055] Specifically, p(B|A,ω f ) can be understood as: given the parameter ω in the initial duration prediction model f and the probability of observing B given input A.

[0056] Specifically, p(B|A, D) can be understood as the probability of observing B given the historical data set D and input A in the initial duration prediction model.

[0057] S2. Train the initial duration prediction model based on a preset training data set and optimize the initial duration prediction model based on a joint loss optimization method to obtain a target duration prediction model. Those skilled in the art know that the preset training data set is a data set pre-set by those skilled in the art according to actual needs, and will not be repeated here.

[0058] Specifically, in the process of optimizing the initial duration prediction model based on the joint loss optimization method, the joint loss function is obtained through the following steps:

[0059] Get G j The corresponding loss function weight N j , N j Meet the following conditions:

[0060] N j =n×exp(U j / T) / (Σ n j＝1 exp(U j / T)), where U j G j The learning rate corresponding to the last training step of , T is the weight of the smoothing factor corresponding to the prediction task.

[0061] According to N j Get the joint loss function Y, which meets the following conditions:

[0062] Y=Σ n j＝1 (N j ×Y j ), Y j G j The loss value.

[0063] Through the above steps, based on the joint loss optimization method, the weights of the loss functions of different prediction tasks in the training process are dynamically adjusted, and the joint loss function is further obtained, which is conducive to improving the stability and performance of model training.

[0064] S3, input the feature value list TZ corresponding to the flight mission to be predicted into the target duration prediction model to obtain the target prediction duration list SC, where TZ = {TZ1, TZ2, ..., TZ i , ..., TZ m}, TZ i For TZ and F i The corresponding eigenvalues, SC = {SC1, SC2, ..., SC j , ..., SC n}, SC j is the predicted duration of the jth target period of the flight mission of the aircraft to be predicted.

[0065] Through the above steps, an initial duration prediction model is established based on a preset task feature label list, a progressive hierarchical extraction framework and a hybrid Bayesian neural network. The initial duration prediction model is trained based on a preset training data set and optimized based on a joint loss optimization method to obtain a target duration prediction model. The feature value list corresponding to the flight mission of the aircraft to be predicted is input into the target duration prediction model to obtain a target prediction duration list. It can be seen that in the process of obtaining the target duration prediction model, the progressive hierarchical extraction is used as a framework, which has a strong feature extraction capability and can simultaneously predict the duration of multiple target time periods. The hybrid Bayesian neural network is introduced, which effectively combines the advantages of prior knowledge and data-driven technology, significantly enhances the robustness and prediction accuracy of the target duration prediction model, and dynamically adjusts the weights of the loss function of different prediction tasks in the training process, which is conducive to improving the stability and performance of the model training. Therefore, predicting the duration based on the target duration prediction model is conducive to improving the efficiency and accuracy of duration prediction.

[0066] In a specific embodiment, the following steps are also included before step S1:

[0067] Get the initial time period ID list AA = {AA (1) , A.A. (2) , ..., AA (ai) , ..., AA (am)}, where AA (ai) is the ID of the initial time period ai, the value of ai ranges from 1 to am, and am is the number of initial time period IDs.

[0068] Specifically, the initial period ID is a unique identifier of the initial period.

[0069] Specifically, the initial period is the time period between any two adjacent preset execution actions.

[0070] Specifically, in a specific embodiment, when the aircraft's flight mission is a flight, the preset execution action can be understood as a series of ground handling actions for the aircraft, passengers, and cargo and mail carried out by the ground handling department to ensure the normal operation of the flight after the aircraft arrives at the port, such as adding wheel blocks, removing wheel blocks, and disembarking passengers.

[0071] Further, in a specific embodiment, when the flight mission of the aircraft is a flight, since there is a short-term continuity and synchronization between the nodes of wheel chock, bridge / passenger elevator docking and cabin door opening, that is, after the wheel chock operation is completed, the bridge / passenger elevator docking is immediately followed, and the cabin door is immediately opened and closed, so they can be combined into a preset execution action; since there is also a short-term continuity and synchronization between the nodes of bridge / passenger elevator evacuation and cabin door closing, that is, after the wheel chock removal operation is completed, the bridge / passenger elevator docking is immediately followed, and the cabin door is immediately opened and closed, so they can be combined into a preset execution action;

[0072] Get AA (ai) Corresponding historical duration list AB (ai) ={AB (ai)(1) , A.B. (ai)(2) , ..., AB (ai)(aj) , ..., AB (ai)(an)}, A B (ai)(aj) For the ajth historical aircraft mission AA (ai) The duration of the corresponding initial period, aj takes the value of a1 to an, an is the number of historical aircraft flight missions, and the historical aircraft flight missions are the aircraft flight missions that operate normally before the current time point.

[0073] Obtaining AA based on Pearson correlation coefficient method (ai) With AA (ae) The corresponding correlation coefficient AC (ai)(ae) , where AA (ae) is the ID of the initial time period ae, the value of ae ranges from 1 to am, AC (ai)(ae) Meet the following conditions:

[0074] Among them, AB (ae)(aj) For the ajth historical aircraft mission AA (ae) The duration of the corresponding initial period.

[0075] When any |AC (ai)(ae) |>AC 0 When ai≠ae, set AA (ai) The corresponding initial period is taken as the target period, AC 0 is the preset correlation coefficient threshold, AC 0 The value range is (0, 1).

[0076] Through the above steps, based on the historical duration list corresponding to the two initial time period IDs, the correlation coefficients corresponding to the two initial time period IDs are obtained, and based on the correlation coefficients corresponding to the two initial time period IDs and the preset correlation coefficient threshold, the initial time period with correlation is determined and the initial time period is used as the target time period to ensure that the target duration prediction model can be used smoothly.

[0077] In a specific embodiment, the evaluation index of the target duration prediction model is obtained by the following steps:

[0078] Construct an initial experimental dataset and preprocess the initial experimental dataset to obtain a key experimental dataset.

[0079] Based on the key experimental data set, the experimental aircraft flight mission ID list AD, the feature value list set AE corresponding to AD, and the actual duration list set AF corresponding to AD are obtained, where AD = {AD (1) , A.D. (2) , ..., A.D. (ak) , ..., A.D. (at)}, A.D. (ak) is the akth experimental aircraft mission ID, ak ranges from 1 to at, at is the number of experimental aircraft missions, and the experimental aircraft mission is the aircraft mission that provides initial experimental data, AE = {AE (1) , A.E. (2) ,……,AE (ak) ,……,AE (at)}, AE (ak) ={AE (ak)1 , A.E. (ak)2 ,……,AE (ak)i ,……,AE (ak)m}, AE (ak) AD (ak) The corresponding eigenvalue list of the experimental aircraft flight mission, AE (ak)i For AE (ak) Medium and F i The corresponding eigenvalue, AF = {AF (1) , A.F. (2) , ..., AF (ak) , ..., AF (at)}, AF (ak) ={AF (ak)1 , A.F. (ak)2 , ..., AF (ak)j , ..., AF (ak)n}, AE (ak) AD (ak) The actual duration of the corresponding experimental aircraft flight missions, AF (ak)j AD (ak) The actual duration of the jth target period in the corresponding experimental aircraft flight mission.

[0080] Specifically, the experimental aircraft flight mission ID is the unique identity of the experimental aircraft flight mission.

[0081] AE(ak) Input into the target duration prediction model to obtain AD (ak) The predicted duration of the jth target period in the corresponding experimental aircraft flight mission AG (ak)j .

[0082] Get the mean absolute error MAE corresponding to the jth target period j , root mean square error RMSE j and R-squared score R 2 j And based on MAE j 、RMSE j and R 2 j The performance of the target duration prediction model is evaluated, where MAE j 、RMSE j and R 2 j Meet the following conditions respectively:

[0083] MAE j =1 / at×Σ at ak＝1 |AG (ak)j -AF (ak)j |RMSE j =(1 / at×Σ at ak＝1 (AG (ak)j -AF (ak)j ) 2 ) 1 / 2 ;

[0084] R 2 j =1-(Σ at ak＝1 (AF (ak)j -AG (ak)j ) 2 / Σ at ak＝1 (AF (ak)j -Σ at ak＝1 AF (ak)j / at) 2 ).

[0085] Specifically, compared with the PLE model and the MMOE model, the target duration prediction model is better in MAE j 、RMSE j and R 2 j The above are significantly better than the PLE model and the MMOE model, as shown in Table 1:

[0086] Table 1: MAE, RMSE, and R2 Comparison Table

[0087]

[0088]

[0089] Through the above steps, both the PLE model and the MMOE model can realize multi-task learning and feature extraction. Furthermore, both the PLE model and the MMOE model can predict the duration of the target time period. However, the experimental results show that the duration predicted by the target duration prediction model is more accurate than that predicted by the PLE model and the MMOE model, and has a higher R-square score, which proves that the target duration prediction model has higher fitting accuracy.

[0090] An embodiment of the present invention also provides a non-transitory computer-readable storage medium, which can be set in an electronic device to store a computer program related to a method in a method embodiment, and the computer program is loaded and executed by the processor to implement the method provided in the above embodiment.

[0091] An embodiment of the present invention further provides an electronic device, comprising: a processor, a memory, and a computer program stored in the memory and executable on the processor, wherein the processor implements the method provided in the above embodiment when executing the computer program.

[0092] An embodiment of the present invention further provides a computer program product, which includes program code. When the program product is run on an electronic device, the program code is used to enable the electronic device to execute the steps of the method according to various exemplary embodiments of the present invention described above in this specification.

[0093] The present invention provides a method, device and medium for predicting the duration of a target time period. The method establishes an initial duration prediction model based on a preset task feature label list, a progressive hierarchical extraction framework and a hybrid Bayesian neural network, trains the initial duration prediction model based on a preset training data set and optimizes the initial duration prediction model based on a joint loss optimization method to obtain a target duration prediction model, inputs a feature value list corresponding to a flight mission of an aircraft to be predicted into the target duration prediction model to obtain a target prediction duration list. It can be seen that the target duration prediction model proposed by the present invention combines the progressive hierarchical extraction framework and the hybrid Bayesian neural network, has a strong feature extraction capability and can simultaneously predict the durations of multiple target time periods, which is beneficial to improving the efficiency and accuracy of duration prediction.

[0094] Although some specific embodiments of the present invention have been described in detail by way of example, it should be understood by those skilled in the art that the above examples are only for illustration, not for limiting the scope of the present invention. It should also be understood by those skilled in the art that various modifications may be made to the embodiments without departing from the scope and spirit of the present invention.

Claims

1. A method for predicting the duration of a target period, characterized in that: The target time period is a time period between adjacent preset execution actions, and the preset execution action is a preparatory action that needs to be performed before the aircraft flight mission is performed. The method includes the following steps: S1. Establish an initial duration prediction model based on the preset task feature label list F, the progressive hierarchical extraction framework and the hybrid Bayesian neural network, where F = {F1, F2, ..., F i , ..., F m }, F i is the preset task feature label of the i-th preset task feature of the aircraft flight mission, the value of i ranges from 1 to m, m is the number of preset task feature labels, and the input of the initial duration prediction model is the feature value list A corresponding to the aircraft flight mission = {A1, A2, ..., A i , ..., A m }, A i For A and F i The output of the initial duration prediction model is the predicted duration list B = {B1, B2, ..., B j , ..., B n }, B j is the predicted duration obtained after the j-th prediction task predicts the j-th target period, j ranges from 1 to n, n is the number of prediction tasks, the initial duration prediction model includes a first feature extraction module and a second feature extraction module, the first feature extraction module is used to extract shallow features, the second feature extraction module is used to extract deep features, the structures of the first feature extraction module and the second feature extraction module are both progressive hierarchical extraction structures, the first feature extraction module and the second feature extraction module are both composed of n first-class networks and several second-class networks, the first-class network is a neural network used to extract specific features of a single prediction task, the second-class network is a neural network used to extract common features of all prediction tasks, and the first-class network and the second-class network are both constructed based on the Bayesian neural network structure; S2. Training the initial duration prediction model based on a preset training data set and optimizing the initial duration prediction model based on a joint loss optimization method to obtain a target duration prediction model; S3, input the feature value list TZ corresponding to the flight mission to be predicted into the target duration prediction model to obtain the target prediction duration list SC, where TZ = {TZ1, TZ2, ..., TZ i , ..., TZ m }, TZ i For TZ and F i The corresponding eigenvalues, SC = {SC1, SC2, ..., SC j , ..., SC n }, SC j is the predicted duration of the jth target period of the flight mission of the aircraft to be predicted.

2. The method for predicting the duration of a target period according to claim 1, characterized in that: The first type of network corresponds one-to-one to the prediction task.

3. The method for predicting the duration of a target period according to claim 1, characterized in that: The initial duration prediction model also includes n task towers.

4. The method for predicting the duration of a target period according to claim 3, characterized in that: The task towers correspond one to one with the prediction tasks.

5. The method for predicting the duration of a target period according to claim 3, characterized in that: The mission tower is built based on a fully connected network structure.

6. The method for predicting the duration of a target period according to claim 3, characterized in that: The first feature extraction module and the second feature extraction module are in communication connection with each task tower.

7. The method for predicting the duration of a target period according to claim 1, characterized in that: In the initial duration prediction model, the self-attention layer is used to obtain the attention weight list set C corresponding to A = {C1, C2, ..., C i , ..., C m } and obtain the intermediate feature vector list E corresponding to A based on C = {E1, E2, ..., E i ,……,E m }, where C i A i The corresponding attention weight list, E i A i The corresponding intermediate feature vector includes the following steps: Get A i The corresponding query vector Q i , key vector K i Sum value vector V i , Q i , K i and V i Meet the following conditions respectively: Q i =W Q ×A i , K i =W K ×A i , V i =W V ×A i , W Q is the weight matrix used to calculate the query vector, W K is the weight matrix used to calculate the key vector, W V is the weight matrix corresponding to the value vector used to calculate; Get C i ={C i1 , C i2 , ..., C iα , ..., C im }, C iα Q i K α The attention weight, K α A α The corresponding key vector, A α For A and F α The corresponding eigenvalue, F α is the preset task feature label of the αth preset task feature of the aircraft flight mission, where α ranges from 1 to m, and C iα Meet the following conditions: C iα =softmax(Q i ×K α T / d k 1 / 2 ), where K α T K α The transposed matrix, d k is the dimension of the key vector; Get E i , where E i Meet the following conditions: E i =∑ m α＝1 (C iα ×V α ), V α A α The corresponding value vector.

8. The method for predicting the duration of a target period according to claim 7, characterized in that: During the operation of the initial duration prediction model, B is obtained through the following steps. j : Get the j-th prediction task G according to E j The corresponding selection matrix H j ; According to A i and H j Get G j Middle A i The corresponding first feature weight R ij , R ij Meet the following conditions: R ij =A i ×H j , the first feature weight is used to indicate the importance of the feature value for the prediction task; According to R ij Get G j The corresponding comprehensive feature weight L j , L j Meet the following conditions: L j =∑ m i＝1 (exp(M(A i )) / ∑ m i＝1 exp(M(A i ))×R ij ), where M(A i ) is the key module for A i The second feature weight obtained after processing, the key module is the module containing the gated recurrent unit and the attention mechanism in the initial duration prediction model; L j Input to G j Get B in the corresponding mission tower j .

9. A non-transitory computer-readable storage medium, characterized in that: The storage medium stores a computer program, which is loaded and executed by a processor to implement the method for predicting the duration of a target time period as described in any one of claims 1 to 8.

10. An electronic device comprising: A processor, a memory, and a computer program stored in the memory and executable on the processor, wherein the processor implements the method for predicting the duration of a target time period as described in any one of claims 1 to 8 when executing the computer program.

Images