Flight ground support data filling model construction method

Abstract

The invention provides a flight ground guarantee data filling model construction method, which comprises the steps of constructing a data filling model, comprising a time sequence position coding module, a topological feature coding module, a mask matrix generation module, a time sequence matrix generation module, a first splicing module, a feature fusion module, a pre-filling matrix generation module, a second splicing module, a filling module, a first updating module, a third splicing module, a reconstruction module, a weighted fusion module and a second updating module. Obtaining an initial time sequence characteristic matrix set R based on the original data in the set time period, and dividing the R into a training set, a verification set and a test set according to a preset proportion; training the initial data filling model by using the training set to obtain a trained data filling model; and verifying the trained data filling model by using the verification set, and testing the final data filling model by using the test set. According to the invention, the data filling accuracy can be improved.

Description

Technical Field

[0001] The present invention relates to the field of data processing, and in particular to a method for constructing a flight ground support data filling model. Background Art

[0002] With the rapid development of civil aviation transportation, the number of take-offs and landings of airport flights and the throughput of passengers are increasing. The contradiction between the increasing number of flights and the limited resources of the airport has gradually become prominent. How to ensure that all flights can operate normally has become a key issue. Flight ground support service refers to the various work processes between the two support nodes of the aircraft from arriving at the airport, entering the parking position to complete the wheel block operation to completing the wheel block removal operation, with aircraft, passengers, baggage and mail as the core, including unloading passengers and cargo, cleaning the aircraft, refueling, replenishment and other ground operations, involving multiple complex tasks and resources, including crew, ground staff, equipment and facilities, etc., which brings great difficulty to accurately predict the flight ground support time; in addition, since flight ground support data is usually obtained by sensor equipment or recorded manually, data loss and anomalies caused by sensor failure or manual recording errors are inevitable, resulting in the current airport collaborative decision-making system still has local data loss and anomalies, inaccurate prediction algorithms, and failure to truly realize the "collaborative decision-making" of stakeholders, etc., thus failing to provide accurate auxiliary decision-making and strong technical support for flight ground support, which also brings many uncertain challenges to downstream tasks such as flight ground support milestone events. Summary of the invention

[0003] In view of the above technical problems, the technical solution adopted by the present invention is:

[0004] According to a first aspect of the present invention, a method for constructing a flight ground support data filling model is provided, wherein the flight ground support data is flight ground support milestone event occurrence time data, and the method comprises the following steps:

[0005] S100, constructing a flight ground support data filling model; the data filling model includes a temporal position coding module, a topological feature coding module, a mask matrix generation module, a time series matrix generation module, a first splicing module, a feature fusion module, a pre-filling matrix generation module, a second splicing module, a filling module, a first update module, a third splicing module, a reconstruction module, a weighted fusion module and a second update module; wherein the temporal position coding module is used to perform position coding on the received feature matrix, obtain the corresponding position coding matrix and input it into the time series matrix module, the topological feature coding module is used to encode the topological features of the received feature matrix, obtain the corresponding topological feature coding matrix and input it into the first splicing module, and the mask matrix generation module is used to perform a mask operation on the received feature matrix, The corresponding mask matrix is ​​obtained and input into the time series matrix generation module, the first splicing module, the second splicing module and the third splicing module respectively. The output of the time series matrix generation module is used as the input of the first splicing module and the pre-filling matrix generation module respectively. The output of the first splicing module is used as the input of the feature fusion module, the output of the feature fusion module is used as the input of the pre-filling matrix generation module, the output of the pre-filling matrix generation module is used as the input of the second splicing module, the output of the second splicing module is used as the input of the filling module, the output of the filling module is used as the input of the first updating module, the output of the first updating module is used as the input of the third splicing module, the output of the third splicing module is used as the input of the reconstruction module, the output of the reconstruction module is used as the input of the weighted fusion module, and the output of the weighted fusion module is used as the input of the second updating module.

[0006] S200, based on the original data in the set time period, an initial time series feature matrix set R is obtained, and R is divided into a training set, a validation set and a test set according to a preset ratio; wherein the original data includes flight attribute data and flight ground support milestone event occurrence time data; R = {R i} i＝1……n , R i is the initial time series feature matrix corresponding to the ith flight, i ranges from 1 to n, n is the number of flights, R i The j-th row vector x in j i ={x jk i} k＝1……(m-1) , x jk i is the representation value corresponding to the k+1th milestone event occurrence time data of the i-th flight at the j-th unit time. The value of j ranges from 1 to d, d is the number of unit time in the set time period, and the value of k ranges from 1 to m-1, m is the number of milestone events in the original data.

[0007] S300, training the flight ground support data filling model using the training set to obtain a trained flight ground support data filling model.

[0008] S400, using the verification set to verify the trained flight ground support data filling model to obtain a corresponding verification result. If the verification result indicates that the trained flight ground support data filling model meets the preset conditions, the trained flight ground support data filling model is used as the final flight ground support data filling model and S500 is executed; if the verification result indicates that the trained flight ground support data filling model does not meet the preset conditions, S300 is executed.

[0009] S500, use the test set to test the final flight ground support data filling model.

[0010] A second aspect of the present invention provides an electronic device, comprising a processor and a memory; the processor is used to execute the steps of the method described in the first aspect of the present invention by calling a program or instruction stored in the memory.

[0011] According to a second aspect of the present invention, there is provided a computer-readable storage medium storing a program or instructions, wherein the program or instructions enable a computer to execute the steps of the method according to the first aspect of the present invention.

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

[0013] The flight ground support data filling model construction method provided by the embodiment of the present invention adopts a combination of a Transformer-based encoder and a linear layer to construct a flight ground support data filling model, wherein, in view of the characteristics of the flight ground support milestone event time data, a topological relationship probability matrix construction method is used, the introduction of the matrix provides a basis for data filling, and the self-attention mechanism is optimized, which enables the constructed model to learn while paying attention to global information, while also better capturing local information, so that the model can fully learn the historical information and future information of the sequence. The present invention can improve the accuracy of filling missing data for flight ground support.

[0014] It should be understood that the contents described in this section are not intended to identify the key or important features of the embodiments of the present invention, nor are they intended to limit the scope of the present invention. Other features of the present invention will become easily understood through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] 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.

[0016] Figure 1 A flowchart of a method for constructing a flight ground support data filling model provided by an embodiment of the present invention;

[0017] Figure 2 A schematic diagram of a flight ground support process provided by an embodiment of the present invention;

[0018] Figure 3 A structural block diagram of a data filling model provided by an embodiment of the present invention;

[0019] Figure 4 and Figure 5 They are schematic diagrams of comparative test results based on original data and filled data designed for the milestone event of opening and closing the cabin door;

[0020] Figure 6 and Figure 7 The following are schematic diagrams of comparative test results based on original data and filled data designed for milestone events in the operation of the corridor bridge. DETAILED DESCRIPTION

[0021] 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.

[0022] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as those commonly understood by those skilled in the art of the present invention. The terms used herein in the specification of the present invention are only for the purpose of describing specific embodiments and are not intended to limit the present invention. The term "and / or" used herein includes any and all combinations of one or more related listed items.

[0023] It should be noted that some exemplary embodiments are described as processes or methods depicted as flow charts. Although the flow charts describe the steps as sequential processes, many of the steps can be implemented in parallel, concurrently or simultaneously. In addition, the order of the steps can be rearranged. The process can be terminated when its operation is completed, but it can also have additional steps not included in the accompanying drawings. The process can correspond to a method, function, procedure, subroutine, subprogram, etc.

[0024] An embodiment of the present invention provides a method for constructing a flight ground support data filling model to construct a data filling module capable of accurately filling in missing data, wherein the flight ground support data is flight ground support milestone event occurrence time data.

[0025] In the embodiment of the present invention, the flight ground support milestone event refers to the support business corresponding to each node in the flight ground support process. In the embodiment of the present invention, the key business of flight ground support is used as the process node, and the flight operation support process can be abstracted as follows: Figure 1 , where the nodes represent specific support services, i.e., milestone events, such as boarding, disembarking, refueling, maintenance, etc., and the arrows represent the execution order of support services. The execution time corresponding to each node is called the time data of the milestone event of flight ground support. For example, the time when the wheel block starts and the time when the wheel block ends are all the time data of the milestone event of flight ground support.

[0026] Further, in an embodiment of the present invention, if Figure 2 As shown, an embodiment of the present invention provides a method for constructing a flight ground support data filling model, which may include the following steps:

[0027] S100, constructing a flight ground support data filling model.

[0028] In the embodiment of the present invention, Figure 3 As shown, the flight ground support data filling model may include a temporal position encoding module 1, a topological feature encoding module 2, a mask matrix generation module 3, a time series matrix generation module 4, a first splicing module 5, a feature fusion module 6, a pre-filling matrix generation module 7, a second splicing module 8, a filling module 9, a first updating module 10, a third splicing module 11, a reconstruction module 12, a weighted fusion module 13 and a second updating module 14.

[0029] Among them, the temporal position encoding module 1 is used to perform position encoding on the received feature matrix X, obtain the corresponding position encoding matrix and input it into the time series matrix generation module 4. The topological feature encoding module 2 is used to encode the topological features of the received feature matrix X, obtain the corresponding topological feature encoding matrix, and input it into the first splicing module 5. The mask matrix generation module 3 is used to perform mask operation on the received feature matrix, obtain the corresponding mask matrix and input it into the time series matrix module 4, the first splicing module 5, the second splicing module 8 and the third splicing module 11 respectively. The output of the time series matrix generation module 4 is respectively used as the input of the first splicing module 5 and the pre-filling matrix generation module 7, the output of the first splicing module 5 is used as the input of the feature fusion module 6, the output of the feature fusion module 6 is used as the input of the pre-filling matrix generation module 7, the output of the pre-filling matrix generation module 7 is used as the input of the second splicing module 8, the output of the second splicing module 8 is used as the input of the filling module 9, the output of the filling module 9 is used as the input of the first updating module 10, the output of the first updating module 10 is used as the input of the third splicing module 11, the output of the third splicing module 11 is used as the input of the reconstruction module 12, the output of the reconstruction module 12 is used as the input of the weighted fusion module 13, and the output of the weighted fusion module 13 is used as the input of the second updating module 14.

[0030] S200, based on the original data in the set time period, an initial time series feature matrix set R is obtained, and R is divided into a training set, a validation set and a test set according to a preset ratio; wherein the original data includes flight attribute data and flight ground support milestone event occurrence time data; R = {R i} i＝1……n , R i is the initial time series feature matrix corresponding to the ith flight, i ranges from 1 to n, n is the number of flights, R i The j-th row vector x in j i ={x jk i} k＝1……(m-1) , x jk i is the representation value corresponding to the k+1th milestone event occurrence time data of the i-th flight at the j-th unit time. The value of j ranges from 1 to d, d is the number of unit time in the set time period, and the value of k ranges from 1 to m-1, m is the number of milestone events in the original data.

[0031] In an embodiment of the present invention, the duration of the set time period can be set based on actual needs, for example, it can be set to a formal multiple of 7 days, such as 7 days, 21 days, 28 days, 35 days, 42 days, 49 days, 56 days, etc. In an exemplary embodiment, the duration of the set time period can be 91 days, and the duration of the unit time can be 1 day.

[0032] In an embodiment of the present invention, flight attribute data may include the flight date, inbound flight number, outbound flight number, aircraft model, aircraft number and aircraft position; flight ground support milestone event occurrence time data may include actual wheel block time, jet bridge time, cabin door opening time, unloading start time, unloading end time, refueling end time, meal start time, meal end time, baggage loading start time, baggage loading end time, cabin door closing time, jet bridge removal time, actual wheel block removal time and other 13 node time data, that is, in the embodiment of the present invention, m=13.

[0033] In this embodiment of the present invention, the data set of the time when each milestone event of flight f occurs on day h is recorded as vector X h f ={x h1 f , x h2 f , ..., x hm f}, where x h1 f represents the last wheel blocking time of flight f on day h, x hm f represents the off-block time of flight f on day h. Thus, based on the data set of the occurrence time of each milestone event of flight f within the set time period, the original matrix of flight f is constructed.

[0034] In the embodiment of the present invention, since the time when the flight ground support milestone event occurs is the actual time, such as 12:30 on August 25, 2024, this data cannot be directly applied. Therefore, the present invention uses the upper block time as the reference point, and the offset obtained by subtracting the remaining 12 node time data from the upper block time is the data stored in the matrix R, that is, x jk i =x0 jk i -xc j i , x0 jk i is the time data of the k+1th milestone event of the i-th flight at the j-th unit time, xc j i It is the time data of the first milestone event of the i-th flight at the j-th unit time.

[0035] In an embodiment of the present invention, an experiment is conducted using real airport flight ground support data collected by an A-CDM system at a hub airport in North China. The original data is collected through various sensors deployed at the airport and manual records. Table 1 shows the data records after the original inbound and outbound flight data are merged and processed by the "flight schedule date + flight number" combined primary key. The main attributes include the flight schedule date, inbound / outbound flight number, scheduled flight arrival and departure time, and a series of flight ground service node times. The time period considered is from October 21 to November 20, 2023, and the flight time every day is from 00:00 in the morning to 24:00 in the evening. A total of 25,544 flights are included, including 12,770 inbound flights and 12,774 outbound flights, with an average of about 851 flights per day.

[0036] Table 1:

[0037]

[0038] In the embodiment of the present invention, the training set, the validation set and the test set are divided in a ratio of 7:1:2.

[0039] S300, training the flight ground support data filling model using the training set to obtain a trained flight ground support data filling model.

[0040] Furthermore, S300 may specifically include:

[0041] S301, use the time series position encoding module 1 to position encode the initial time series feature matrix currently input, obtain the corresponding time series position encoding matrix, use it as the current time series position encoding matrix and input it to the time series matrix generation module 4; use the topological feature encoding module 2 to obtain the topological feature encoding matrix corresponding to the initial time series feature matrix currently input, use it as the current topological feature encoding matrix and input it to the first splicing module 5; use the mask matrix generation module 3 to randomly generate a mask matrix of the initial time series feature matrix currently input, use it as the current mask matrix and input it to the first splicing module 5, the second splicing module 8 and the third splicing module 11 respectively; wherein, the topological feature encoding matrix corresponding to the initial time series feature matrix currently input is used to characterize the correlation between the occurrence time data of any two flight ground support milestone events in the initial time series feature matrix currently input.

[0042] In the embodiment of the present invention, the temporal position encoding module 1 performs position encoding on the matrix in the following manner:

[0043] For the k1th odd bit of the data in the jth row vector in the matrix, the encoding result is sin(j / 10000 (2k1 / (m-1))), for the k2th even-numbered bit of the jth row vector in the matrix, the encoding result is cos(j / 10000 (2k2 / (m-1)) ).

[0044] In this embodiment of the present invention, the current mask matrix is ​​obtained by the following steps:

[0045] S10, randomly set the elements in the current input initial time series feature matrix that are not null values ​​to null values, and obtain an initial time series feature matrix with random missing values.

[0046] S20, constructing a mask matrix based on the initial time series feature matrix with random missing values ​​to obtain a current mask matrix, wherein the value of each element in the current mask matrix is ​​a mask identifier determined based on the data at the corresponding position of the initial time series feature matrix with random missing values, if the value of an element of the initial time series feature matrix with random missing values ​​is a null value and is not a randomly generated null value, that is, an original missing value, then the value of the corresponding position in the current mask matrix is ​​a first mask identifier, if the value of an element of the initial time series feature matrix with random missing values ​​is not a null value, that is, not missing, then the value of the corresponding position in the current mask matrix is ​​a second mask identifier, if the value of an element in the initial time series feature matrix with random missing values ​​is a randomly generated null value, that is, an active missing value, then the value of the corresponding position in the current mask matrix is ​​a third mask identifier.

[0047] In the embodiment of the present invention, the first mask identifier may be -1, the second mask identifier may be 1, and the third mask identifier may be 0.

[0048] In the embodiment of the present invention, the originally missing part refers to the data of the original position, that is, there is no data at the position, and the non-missing part refers to the data of the original position, that is, the corresponding position. Since the real data of the original missing cannot be known during the model training process, in order to verify the effectiveness of the method of the present invention, active missing is introduced. Active missing means that there is data at the original position. In order to verify the accuracy of the filling model, it is artificially constructed to be missing, so that its value in the input matrix is ​​0.

[0049] In an embodiment of the present invention, in view of the characteristics of the topological dependency relationship between the occurrence times of flight ground support milestone events, the present invention proposes a topological relationship probability matrix construction method, by modeling the flight ground support network into a graph structure, using the adjacency matrix of the graph to represent the dependency relationship between different milestone events, the flight ground support network can be modeled as a graph G = (V, E, A). Among them, V is a set of a series of milestone events, N = |V| represents the size of the set, E represents the edge formed between each milestone event in the network, that is, the dependency relationship, and A represents a topological relationship matrix, the size of which is (m-1) × (m-1). In an embodiment of the present invention, the degree of correlation between any two flight ground support milestone event occurrence time data in the currently input initial time series feature matrix is ​​the degree of correlation between the corresponding column vectors in the matrix, which can be represented by a correlation coefficient. In an embodiment of the present invention, the correlation coefficient of the column vectors can be obtained by using methods such as the Pearson correlation coefficient, the Spearman correlation coefficient, and the Kendall correlation coefficient. Preferably, the Pearson correlation coefficient is used to obtain it.

[0050] S302, using the time series matrix generation module 4 to add the current input initial time series feature matrix and the current time series position coding matrix, the addition result is multiplied by the current mask matrix to obtain the time series matrix corresponding to the current input initial time series feature matrix, as the current time series matrix and input into the first splicing module 5 and the pre-filling matrix generation module 7 respectively.

[0051] In an embodiment of the present invention, the initial time series feature matrix and the current time series position coding matrix are added to mark the position information, and then the addition result is multiplied by the mask matrix to construct active missing data and mark the data missing information, thereby obtaining a time series matrix with position information and mask information.

[0052] S303, using the first splicing module 5 to splice the current time series matrix, the current mask matrix and the current topological feature encoding matrix into a three-channel two-dimensional matrix, and using the feature fusion module 6 to perform feature fusion on the two-dimensional matrix to obtain the corresponding fusion result as the current fusion feature matrix and input it into the pre-filled matrix generation module 7.

[0053] In an embodiment of the present invention, the feature fusion module can be a 1×1 convolution layer. By splicing the current time series matrix, the current mask matrix and the current topological feature encoding matrix into a three-channel two-dimensional matrix and inputting it into the 1×1 convolution layer for feature extraction, the information of the spliced ​​matrix on the three channels can be integrated to complete the dimensionality reduction operation, and the dependency information between the time dimension and the features can be effectively extracted, so that the valid values ​​in the fusion sequence are used to update the missing values.

[0054] S304, using the pre-filling matrix generation module 7 to fuse the current time series matrix and the current fusion feature matrix, obtain the corresponding fusion result as the current pre-filling matrix and input it into the second splicing module 8.

[0055] In the embodiment of the present invention, the current pre-filled matrix X is obtained c The following conditions must be met:

[0056] X c =W×(I n +(UX n +(VX n T ) T )+B);

[0057] Where W, U and V represent the learnable weight parameters of the pre-filled matrix generation module, B is the bias parameter of the pre-filled matrix generation module, T represents the transpose, and I n is the current fusion feature matrix, X n is the current time series matrix, UX n Indicates the impact on the time series dimension, VX n T Indicates the impact between different milestone event nodes.

[0058] S305 , using the second splicing module 8 to splice the current pre-filling matrix and the current mask matrix, and inputting the splicing result into the filling module 9 for filling processing, obtaining the corresponding filling result as the current filling matrix and inputting it into the first updating module 10 .

[0059] In an embodiment of the present invention, the padding module 9 may include a convolutional self-attention layer, a feedforward fully connected layer and a linear layer connected in sequence.

[0060] The traditional self-attention mechanism has certain advantages in capturing global information, but has certain limitations in capturing local information. Combined with the characteristics of the flight ground support milestone event data filling task, the present invention proposes an improved convolutional self-attention mechanism. Since the missing values ​​of the active missing positions are initialized to 0, when the data missing rate is high, such as 70% missing, the extracted information will have a high degree of similarity. In order to avoid this phenomenon affecting the calculation of weight distribution in the self-attention layer, a one-dimensional convolutional layer is added to better capture local feature information, and then the output of the one-dimensional convolutional layer is linearly transformed to obtain Q, K and V, that is, the convolutional self-attention layer includes sequentially connected convolutional layers and self-attention mechanism layers, wherein the convolutional layer is a 1×1 convolutional layer and the number of heads of the self-attention mechanism layer is single.

[0061] Among them, the output of the convolutional layer of the convolutional self-attention layer is C = (c 1, ..., c j , ..., c d ), c j is the convolution feature corresponding to the jth unit time in the splicing result obtained by the second splicing module 8, c j =w j (F jk ) k＝1……(m-1) , F jk is the feature corresponding to the data of the occurrence time of the kth milestone event in the jth unit time in the splicing result obtained by the second splicing module 8, w j is the weight matrix corresponding to the jth unit time in the splicing result obtained by the second splicing module 8. The output of the self-attention mechanism layer Out = Softamax (QK T / (m-1) 1 / 2 )V, where Q is the query vector, K is the key vector, and V is the value vector. Q = (q 1 , ..., q j , ..., q d ), K=(k 1 , ..., k j , ..., k d ), V=(v 1 , ..., v j , ..., v d ), the jth query value q j =c j × q +b q , the jth key value k j =c j × k +b k , the jth value v j =v j × v +b v , w q 、w k and w v are key weight, query weight and value weight respectively, b q , b k and b v They are key bias, query bias and value bias respectively, and Softamax() represents the normalized exponential function.

[0062] In the embodiment of the present invention, the feed-forward fully connected layer is a fully connected network with two linear layers, and the calculation formula is as follows:

[0063] FFN(X 1 ) = ReLU(ReLU(X 1 W 1 +b 1 )W2 +b 2 );

[0064] Among them, FFN(X 1 ) is the output of the feedforward neural network, X 1 is the input of the feedforward neural network, that is, the output of the convolutional self-attention layer of the padding module, W 1 , W 2 is the learnable weight parameter of the feedforward fully connected layer, b 1 , b 2 is the bias parameter of the feedforward fully connected layer, and ReLU() is the activation function.

[0065] The output of the linear layer is the current padding matrix X f1 For: X f1 =FFN(X 1 )W 3 +b 3 ; W 3 is the learnable weight parameter of the linear layer, b 3 is the bias parameter of the linear layer.

[0066] S306 , the first updating module updates the current time series feature matrix using the current filling matrix to obtain a first updated time series matrix and inputs it into the third splicing module 11 .

[0067] In an embodiment of the present invention, the first updating module is used to replace the elements with the third mask identifier in the current time series matrix with the elements at the corresponding positions in the current filling matrix, that is, to replace the elements with the mask identifier of 0 in the current time series matrix with the elements at the corresponding positions in the current filling matrix, and obtain the current time series matrix after filling as the current filling matrix.

[0068] S307, using the third splicing module 11 to splice the current padding matrix and the current mask matrix, and input the splicing result into the reconstruction module 12 for reconstruction processing, to obtain the corresponding reconstruction result as the current reconstruction matrix and input it into the weighted fusion module.

[0069] In the embodiment of the present invention, the reconstruction module 12 may include a convolutional self-attention layer, a feedforward fully connected layer and a linear layer unit, wherein the convolutional self-attention layer and the feedforward fully connected layer of the reconstruction module have the same structure as the convolutional self-attention layer and the feedforward fully connected layer of the filling module. The linear layer unit is composed of two linear layers and an activation function, that is, the output X of the linear layer unit f2 Meet the following conditions: X f2 =ReLU(ReLU(Z 2 W a +b a )W b +bb ), Z 2 is the output of the feedforward fully connected layer of the reconstruction module, Z 2 =FFN(X 2 ), X 2 is the output of the convolutional self-attention layer of the reconstruction module, W a , W b is the learnable weight parameter of the reconstruction module, b a and b b is the bias parameter of the reconstruction module.

[0070] S308 , using the weighted fusion module 13 to perform weighted fusion on the current filling matrix and the current reconstruction matrix, to obtain a corresponding weighted fusion result as the current weighted filling matrix and input it into the second updating module 14 .

[0071] In this embodiment of the present invention, the current weighted padding matrix satisfies the following conditions:

[0072] Xw = λ × Xf + (1-λ) × Xr;

[0073] Among them, Xw is the current weighted padding matrix, Xf is the current padding matrix, Xr is the current reconstruction matrix, and λ is the padding weight, which is determined based on the weights of the convolutional self-attention layers of the padding module and the reconstruction module.

[0074] Furthermore, λ satisfies the following conditions:

[0075] λ=(α+β) / 2, where α=Softmax(h1 / (h1+h2)), h1 is the weight of the self-attention layer of the filling module, h2 is the weight of the self-attention layer of the reconstruction module, h1 and h2 are the values ​​obtained during the model training process, which are values ​​between 0 and 1. Softmax() is the activation function, and β is the preset weight, which is a value between 0 and 1 and can be adjusted based on the prediction effect of the model.

[0076] In the present invention, the output of the filling module and the output of the reconstruction module are weighted as the current weighted filling matrix, which can more effectively integrate the context information, thereby making the final result more accurate.

[0077] S309, the second updating module uses the current weighted filling matrix to update the current time series feature matrix to obtain a second updated time series matrix as the current filling result.

[0078] In the embodiment of the present invention, the second updating module 14 is used to replace the elements with the third mask identifier in the current time series matrix with the elements at the corresponding positions in the current weighted filling matrix to obtain a second updated time series matrix as the current filling result.

[0079] S310, based on the current filling result and the corresponding real result, determine whether the current flight ground support data filling model meets the preset convergence condition. If so, use the current flight ground support data filling model as the final data filling model. Otherwise, update the parameters of the current flight ground support data filling model and use the next batch of training data as the current input initial time series feature matrix to execute S301.

[0080] In an embodiment of the present invention, the preset convergence condition can be set based on actual conditions, for example, the number of iterations reaches a set iteration threshold or the loss function value of the model is less than or equal to a set loss threshold and remains unchanged within a set time period.

[0081] In the embodiment of the present invention, the true result of the current filling result is the true data before the mask. The loss function value of the model can be the mean absolute error (MAE) and the root mean square error (RMSE) of the model.

[0082] In the embodiment of the present invention, each batch of training data may be an initial time series feature matrix of a flight.

[0083] S400, using the verification set to verify the trained flight ground support data filling model to obtain a corresponding verification result. If the verification result indicates that the trained flight ground support data filling model meets the preset conditions, the trained flight ground support data filling model is used as the final data filling model and S400 is executed; if the verification result indicates that the trained flight ground support data filling model does not meet the preset conditions, S300 is executed.

[0084] In the embodiment of the present invention, the preset conditions can be set based on actual needs, for example, the mean absolute error and the root mean square error of the model are both less than the preset error.

[0085] Those skilled in the art know that the process of using the validation set to validate the trained flight ground support data filling model may be the same as the process of using the training set to train the model.

[0086] S400, use the test set to test the final flight ground support data filling model.

[0087] In an embodiment of the present invention, the process of testing the final flight ground support data filling model using the test set can refer to the process of training the model using the training set. The difference is that during the test process, a mask matrix is ​​generated by randomly masking 20%, 30% and 40% of the data in the input matrix.

[0088] In actual application scenarios, the data filling model constructed by the flight ground support data filling model construction method provided by the embodiment of the present invention can input the feature matrix corresponding to the data to be filled into the data filling model. The processing process of the data filling model is basically the same as the process of filling the data in steps S301 to S309. The difference is that the mask matrix generation module will not construct active missing for the input feature matrix, but generates a mask matrix based on the original missing. The subsequent update module also updates the original missing data to obtain the final filling result.

[0089] The present invention uses the same data set to compare the filling effect of the data filling model constructed by the flight ground support data filling model construction method provided by the present invention with the existing data filling model. The comparison results are shown in Table 2 below:

[0090] Table 2:

[0091] Model MAE RMSE ARIMA 4.431 6.793 SVR 3.774 5.981 GRU 2.867 4.195 STGCN 2.809 4.106 GWNet 2.793 4.072 GMAN 2.546 4.019 SNFM-GraphSAGE 2.165 3.804 Model of the present invention 2.013 3.725

[0092] Among them, ARIMA represents the autoregressive moving average model, SVR represents the support vector machine, GRU represents the gated recurrent neural network, STGCN represents the spatiotemporal graph convolutional network, GWNet is the model used in the literature (Graph WaveNet for Deep Spatial-Temporal Graph Modeling), GMAN represents the graph multi-attention network, and SNFM-GraphSAGE represents the graph sample aggregation embedding network. It can be seen from Table 2 that the data filling model obtained by the present invention is superior to the existing data filling model in terms of MAE and RMSE, indicating that the data filling model constructed by the flight ground support data filling model construction method provided by the present invention has high prediction accuracy.

[0093] In addition, in order to verify the advantages of the padded data proposed in the method of the present invention in solving downstream prediction tasks, a comparative experiment based on the original data and the padded data was designed for the two milestone events of cabin door opening and closing and jet bridge operation. The comparative experimental results are shown in Figure 2. Figures 4 to 7 As shown, Figure 4 The prediction result before filling in the cabin door opening and closing event, Figure 5 The prediction result after filling in the cabin door opening and closing events, Figure 6 This is the prediction result before the bridge operation event is filled. Figure 7 This is the prediction result after the bridge operation event is filled. The horizontal axis is the time slice and the vertical axis is the time (unit: minute). Figures 4 to 7 It can be seen that the accuracy of downstream prediction tasks using the data padded by the padded model provided by the present invention is higher.

[0094] In summary, the flight ground support data filling model construction method provided by the embodiment of the present invention is improved based on the Transformer idea, and can fully learn the historical information and future information of the sequence. Compared with the original method, the improved self-attention mechanism can better capture local information while paying attention to global information; according to the characteristics of the flight ground support milestone event moment data, a topological relationship probability matrix construction method is proposed, and the introduction of this matrix provides a basis for data filling. The experimental results of the method proposed by the present invention on the same filling task show that compared with the more advanced graph multi-attention network model and graph sample aggregation embedding network model, the mean absolute error is reduced by 0.533 and 0.152 respectively. Therefore, the present invention can effectively improve the accuracy of filling missing data for flight ground support, thereby better supporting the development of downstream prediction tasks.

[0095] An embodiment of the present invention also provides an electronic device, comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are configured to execute the method described in the embodiment of the present invention.

[0096] The embodiment of the present invention further provides a computer-readable storage medium storing computer-executable instructions, wherein the computer instructions are used to execute the method described in the embodiment of the present invention.

[0097] It should be understood that the various forms of processes shown above can be used to reorder, add or delete steps. For example, the steps described in the present invention can be executed in parallel, sequentially or in different orders, as long as the desired results of the technical solution disclosed in the present invention can be achieved, and this document does not limit this.

[0098] The above specific implementations do not constitute a limitation on the protection scope of the present invention. It should be understood by those skilled in the art that various modifications, combinations, sub-combinations and substitutions can be made according to design requirements and other factors. Any modification, equivalent substitution and the like made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for constructing a flight ground support data filling model, characterized in that: The flight ground support data is flight ground support milestone event occurrence time data, and the method comprises the following steps: S100, constructing a flight ground support data filling model; the data filling model includes a temporal position coding module, a topological feature coding module, a mask matrix generation module, a time series matrix generation module, a first splicing module, a feature fusion module, a pre-filling matrix generation module, a second splicing module, a filling module, a first update module, a third splicing module, a reconstruction module, a weighted fusion module and a second update module; wherein the temporal position coding module is used to perform position coding on the received feature matrix, obtain the corresponding position coding matrix and input it into the time series matrix module, the topological feature coding module is used to encode the topological features of the received feature matrix, obtain the corresponding topological feature coding matrix and input it into the first splicing module, and the mask matrix generation module is used to perform a mask operation on the received feature matrix, Obtain the corresponding mask matrix and input it into the time series matrix generation module, the first splicing module, the second splicing module and the third splicing module respectively, the output of the time series matrix generation module is used as the input of the first splicing module and the pre-filling matrix generation module respectively, the output of the first splicing module is used as the input of the feature fusion module, the output of the feature fusion module is used as the input of the pre-filling matrix generation module, the output of the pre-filling matrix generation module is used as the input of the second splicing module, the output of the second splicing module is used as the input of the filling module, the output of the filling module is used as the input of the first updating module, the output of the first updating module is used as the input of the third splicing module, the output of the third splicing module is used as the input of the reconstruction module, the output of the reconstruction module is used as the input of the weighted fusion module, and the output of the weighted fusion module is used as the input of the second updating module; S200, based on the original data in the set time period, an initial time series feature matrix set R is obtained, and R is divided into a training set, a validation set and a test set according to a preset ratio; wherein the original data includes flight attribute data and flight ground support milestone event occurrence time data; R = {R i } i＝1……n , R i is the initial time series feature matrix corresponding to the ith flight, i ranges from 1 to n, n is the number of flights, R i The j-th row vector x in j i ={x jk i } k＝1……(m-1) , x jk i is the representation value corresponding to the k+1th milestone event occurrence time data of the i-th flight at the j-th unit time, j ranges from 1 to d, d is the number of unit time in the set time period, k ranges from 1 to m-1, m is the number of milestone events in the original data; S300, training the flight ground support data filling model using the training set to obtain a trained flight ground support data filling model; S400, using the validation set to validate the trained flight ground support data filling model, and obtain a corresponding validation result. If the validation result indicates that the trained flight ground support data filling model meets the preset conditions, the trained flight ground support data filling model is used as the final flight ground support data filling model, and S500 is executed; if the validation result indicates that the trained flight ground support data filling model does not meet the preset conditions, S300 is executed; S500, use the test set to test the final flight ground support data filling model.

2. The method for constructing a flight ground support data filling model according to claim 1, characterized in that: S300 specifically includes: S301, using the temporal position encoding module to position encode the initial temporal feature matrix currently input, obtain the corresponding temporal position encoding matrix, as the current temporal position encoding matrix and input it into the time series matrix generation module; using the topological feature encoding module to obtain the topological feature encoding matrix corresponding to the initial temporal feature matrix currently input, as the current topological feature encoding matrix and input it into the first splicing module; using the mask matrix generation module to randomly generate the mask matrix of the initial temporal feature matrix currently input, as the current mask matrix and input it into the first splicing module, the second splicing module and the third splicing module respectively; wherein, the topological feature encoding matrix corresponding to the initial temporal feature matrix currently input is used to characterize the correlation between the occurrence time data of any two flight ground support milestone events in the initial temporal feature matrix currently input; S302, using the time series matrix generation module to add the current input initial time series feature matrix and the current time series position coding matrix to obtain the addition result and multiply it by the current mask matrix to obtain the time series matrix corresponding to the current input initial time series feature matrix as the current time series matrix and input it into the first splicing module and the pre-filling matrix generation module respectively; S303, using the first splicing module to splice the current time series matrix, the current mask matrix and the current topological feature encoding matrix into a three-channel two-dimensional matrix, and using the feature fusion module to perform feature fusion on the two-dimensional matrix to obtain a corresponding fusion result as the current fusion feature matrix and input it into the pre-filling matrix generation module; S304, using the pre-filling matrix generation module to fuse the current time series matrix and the current fusion feature matrix to obtain a corresponding fusion result as the current pre-filling matrix and input it into the second splicing module; S305, using the second splicing module to splice the current pre-filling matrix and the current mask matrix, and input the splicing result into the filling module for filling processing, to obtain the corresponding filling result as the current filling matrix and input it into the first updating module; S306, the first updating module updates the current time series feature matrix using the current filling matrix to obtain a first updated time series matrix and inputs it into the third splicing module; S307, using the third splicing module to splice the current padding matrix and the current mask matrix, and input the splicing result into the reconstruction module for reconstruction processing, to obtain the corresponding reconstruction result as the current reconstruction matrix and input it into the weighted fusion module; S308, using a weighted fusion module to perform weighted fusion on the current filling matrix and the current reconstruction matrix to obtain a corresponding weighted fusion result as the current weighted filling matrix and input it into the second updating module; S309, the second updating module updates the current time series feature matrix using the current weighted filling matrix to obtain a second updated time series matrix as the current filling result; S310, based on the current filling result and the corresponding real result, determine whether the current flight ground support data filling model meets the preset convergence condition. If so, use the current flight ground support data filling model as the final flight ground support data filling model. Otherwise, update the parameters of the current flight ground support data filling model and use the next batch of training data as the current input initial time series feature matrix to execute S301.

3. The method for constructing a flight ground support data filling model according to claim 2, characterized in that: The current mask matrix is ​​obtained by the following steps: S10, randomly setting the elements in the currently input initial time series feature matrix that are not null values ​​to null values, to obtain an initial time series feature matrix with random missing values; S20, constructing a mask matrix based on the initial time series feature matrix with random missing values ​​to obtain a current mask matrix, wherein the value of each element in the current mask matrix is ​​a mask identifier determined based on the data at the corresponding position of the initial time series feature matrix with random missing values, if the value of an element of the initial time series feature matrix with random missing values ​​is a null value and is not a randomly generated null value, then the value of the corresponding position in the current mask matrix is ​​a first mask identifier, if the value of an element of the initial time series feature matrix with random missing values ​​is not a null value, then the value of the corresponding position in the current mask matrix is ​​a second mask identifier, if the value of an element of the initial time series feature matrix with random missing values ​​is a randomly generated null value, then the value of the corresponding position in the current mask matrix is ​​a third mask identifier.

4. The method for constructing a flight ground support data filling model according to claim 3, characterized in that: The filling module includes a convolutional self-attention layer, a feedforward fully connected layer and a linear layer connected in sequence, and the convolutional self-attention layer includes a convolutional layer and a self-attention mechanism layer connected in sequence, wherein the first update module is used to replace the elements with the third mask identifier in the current time series matrix with the elements in the corresponding position in the current filling matrix to obtain the current time series matrix after filling as the current filling matrix.

5. The method for constructing a flight ground support data filling model according to claim 3, characterized in that: The reconstruction module includes a convolutional self-attention layer, a feedforward fully connected layer and a linear layer unit, and the linear layer unit is composed of two linear layers and an activation function.

6. The method for constructing a flight ground support data filling model according to claim 3, characterized in that: The current weighted padding matrix satisfies the following conditions: Xw = λ × Xf + (1-λ) × Xr; Where Xw is the current weighted padding matrix, Xf is the current padding matrix, Xr is the current reconstruction matrix, and λ is the padding weight, which is determined based on the weights of the convolutional self-attention layers of the padding module and the reconstruction module. The second updating module is used to replace the elements with the third mask identifier in the current time series matrix with the elements at the corresponding positions in the current weighted filling matrix to obtain a second updated time series matrix as the current filling result.

7. The method for constructing a flight ground support data filling model according to claim 6, characterized in that: λ satisfies the following conditions: λ=(α+β) / 2, where α=Softmax(h1 / (h1+h2)), h1 is the weight of the self-attention layer of the padding module, h2 is the weight of the self-attention layer of the reconstruction module, Softmax() is the activation function, and β is the preset weight.

8. The method for constructing a flight ground support data filling model according to claim 1, characterized in that: x jk i =x0 jk i -xc j i , x0 jk i is the time data of the k+1th milestone event of the i-th flight at the j-th unit time, xc j i It is the time data of the first milestone event of the i-th flight at the j-th unit time.

Images