Method and device for constructing flight wheel block removal time prediction model and medium

Abstract

The invention relates to the field of computer technology application, and discloses a flight wheel block removal time prediction model construction method, equipment and a medium, and the model constructed by the method integrates a frequency enhancement channel attention module based on discrete cosine transform, a time convolution network module and a window mixed sparse architecture. The model solves the problem of an actual flight ground guarantee process through a frequency enhancement channel attention mechanism, processes a Gibbs phenomenon possibly occurring in a time sequence by using Fourier transform, and extracts features of adjacent time slices of a test point by using an expansion convolution technology in a time convolution network. Time slice features are fused into a self-attention mechanism, the dependency relationship between long-distance time steps can be captured, then the processed time features are put into a window mixed sparse architecture, and the window mixed sparse architecture can extract and calculate the dominant time slice features while capturing the long-term dependency relationship, so that the time slice features are extracted and calculated, and the time slice features are extracted and calculated. The prediction effect can be improved while the operation complexity is reduced.

Description

Technical Field

[0001] The present invention relates to the field of computer technology applications, and in particular to a method, device and medium for constructing a flight off-block time prediction model. Background Art

[0002] According to the "Statistical Bulletin on the Development of the Civil Aviation Industry in 2023" issued by the Civil Aviation Administration of China on May 30, 2024, my country's civil aviation industry has achieved very significant development. In 2023, the industry will complete 619.5764 million passenger trips, an increase of 146.1% over last year. The intensive operation of flights poses a severe challenge to the safety and efficient management of airports. In February of the same year, the "Technical Specifications for Airport Collaborative Decision Making System" issued by the Civil Aviation Administration of China clearly stated that the A-CDM (Airport Collaborative Decision Making) system adopts a milestone management model, which effectively monitors flights by setting milestone nodes and optimizes ground support by sharing information. As an important link that directly affects the subsequent scheduling of flights under the A-CDM system mechanism, the actual value of the flight off-block time will be directly related to the continuity of the entire ground support process and the effective allocation of airport resources. However, in actual operation, how to accurately predict the flight off-block time, a key milestone node for flight ground support, is still an important problem in the A-CDM system. Summary of the invention

[0003] In view of the above technical problems, the technical solution adopted by the present invention is: According to a first aspect of the present invention, a method for constructing a flight off-block time prediction model is provided, the method comprising the following steps: S100, constructing an initial flight off-block time prediction model; the model includes: a first processing model and a second processing model, the first processing model includes a first feature encoding module, a first discrete cosine transform-based frequency enhancement channel attention module, a first time convolutional network module and an encoder connected in sequence, the second processing model includes a second feature encoding module, a second discrete cosine transform-based frequency enhancement channel attention module, a second time convolutional network module, a decoder and an output prediction module connected in sequence, wherein the encoder is connected to the decoder, the first discrete cosine transform-based frequency enhancement channel attention module is also connected to the encoder, and the second discrete cosine transform-based frequency enhancement channel attention module is also connected to the decoder.

[0004] S200, obtain a sequence sample set D, and divide the sequence sample set into a training set and a test set; D = {D1, D2, ..., D i ,……,D n}; D iis the i-th time series in D, i ranges from 1 to n, and n is the length of the time series; D i =(D i1 , D i2 , ..., D ij , ..., D im ), D ij D i The occurrence time of the jth flight ground support key operation node in , where j ranges from 1 to m, and m is the number of flight ground support key operation nodes.

[0005] S300: Train the initial flight off-block time prediction model using the training set to obtain a trained flight off-block time prediction model.

[0006] S400, using the test set to test the trained flight off-block time prediction model.

[0007] According to a second aspect of the present invention, there is provided 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.

[0008] 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.

[0009] The present invention has at least the following beneficial effects: A method for constructing a flight off-block time prediction model provided by an embodiment of the present invention, wherein the model constructed by the method integrates a frequency-enhanced channel attention module based on discrete cosine transform, a temporal convolutional network module and a window mixed sparse architecture. The model uses a frequency-enhanced channel attention mechanism to solve the actual flight ground support process and uses Fourier transform to process the Gibbs phenomenon that may appear in the time series, and uses the dilated convolution technology in the temporal convolutional network to extract the features of adjacent time segments of the test point, and integrates the time segment features into the self-attention mechanism, which can capture the dependencies between long-distance time steps, and then put the processed time features into the window mixed sparse architecture. The window mixed sparse architecture can extract and calculate the dominant time segment features while capturing long-term dependencies, and can improve the prediction effect while reducing the computational complexity.

[0010] 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

[0011] 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.

[0012] Figure 1 This is a schematic diagram of the flight ground support process; Figure 2 A flowchart of a method for constructing a flight off-block time prediction model provided by an embodiment of the present invention; Figure 3 A structural block diagram of a flight off-block time prediction model provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0013] 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.

[0014] 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.

[0015] 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.

[0016] Generally, accurate prediction of flight block removal time is achieved based on the time nodes of each operation in the flight ground support process. Specifically, the support process is a serial and parallel service process of each operation between the two support nodes, from the start of block loading to the end of block removal. It can be summarized as four parallel operation processes: maintenance inspection, fuel filling, cabin service, and cargo hold service. The cabin service work includes four sub-parallel operation processes: garbage cleaning, sewage treatment, disinfection, and meal supply. Standardized and reasonable flight ground support process such as Figure 1 As shown in the figure, each operation is carried out interactively. Not only are the serial and parallel relationships between the time nodes closely connected, but the operations are also carried out interactively according to the logical relationship under the support of professionals and special equipment, and finally a complete relationship diagram between the operations of flight ground support is formed. By collecting and analyzing the time nodes of these operations, a prediction model is constructed to provide data support for the prediction of off-block time.

[0017] However, when using the time nodes of ground support to predict the earliest available block removal time, due to the practical requirements of serial and parallel rules, the support operation nodes are always discrete. This discrete time node may itself cause the Gibbs phenomenon when performing time series analysis, that is, overshoot or oscillation occurs at the connection point of flight operation time, resulting in delays or even interruptions in other links. Especially during the peak of support, the probability of sudden equipment failure, uncontrollable delays, and additional service needs is greatly increased, and the discrete distribution between nodes is more obvious. This phenomenon may lead to an imbalance in the time coordination between different operations in actual operation, thereby affecting the overall ground support efficiency and flight punctuality.

[0018] The discontinuity, periodicity, emergencies or spikes between the time nodes of various operations of flight ground support are similar to the characteristics of signals in time series. In order to effectively extract and process various characteristics of various time signals and ensure the effectiveness of the design model, the mainstream approach is to mine the frequency information implicit in the real data set, but the prediction results of the mainstream time series prediction model are unsatisfactory. A very important reason is that the model designed by them still has deficiencies and shortcomings when mining the frequency domain information in the real data set. Most scholars mainly use the frequency mining method based on Fourier transform to obtain frequency information and use inverse Fourier transform to reconstruct time information. However, when such a time series is Fourier transformed or truncated to a limited frequency range, the reconstructed signal will overshoot and oscillate near the discontinuity point, which produces the Gibbs phenomenon. This phenomenon will also cause another unnecessary round of inverse operations in the network, which not only increases the computational complexity but also wastes resources. Similarly, the fluctuation phenomenon of the time node of flight ground support will affect the prediction accuracy of the off-block time. Therefore, it is necessary to mitigate the negative impact of the Gibbs phenomenon to ensure the accuracy of the off-block time prediction and the efficient operation of the support process.

[0019] In addition, when calculating with the traditional self-attention mechanism, the data at each time point will be involved in the calculation, which will bring high computational complexity, especially when the sequence length becomes longer, the amount of calculation will increase exponentially. But in fact, not all time points have a significant impact on the results, only a few key points are more important. In practical applications, the observation points closer to the prediction point have a greater impact on the results and are more important, while the observation points far away from the prediction point contribute less; in addition to the distance factor, the importance of the actual observation points cannot be ignored, and it is also necessary to select observation points that are more important than the prediction points. Therefore, how to effectively use the two types of observation points to reduce the amount of calculation while improving the prediction effect is also an extremely important issue.

[0020] In order to solve the problems of correlation between operation rules, Gibbs phenomenon in time series analysis, and high computational complexity and low computational speed caused by the multi-head attention mechanism in the original Transformer model when processing long sequences, the embodiment of the present invention provides a method for constructing a flight off-block time prediction model, such as Figure 2 As shown, the method comprises the following steps: S100, constructing an initial flight off-block time prediction model.

[0021] In the embodiment of the present invention, Figure 3 As shown, the model includes: a first processing model and a second processing model, the first processing model includes a first feature encoding module 10, a first frequency enhanced channel attention module 11 based on discrete cosine transform, a first time convolutional network module 12 and an encoder 13 connected in sequence, the second processing model includes a second feature encoding module 20, a second frequency enhanced channel attention module 21 based on discrete cosine transform, a second time convolutional network module 22, a decoder 23 and an output prediction module 24 connected in sequence, wherein the encoder is connected to the decoder, the first frequency enhanced channel attention module based on discrete cosine transform is also connected to the encoder, and the second frequency enhanced channel attention module based on discrete cosine transform is also connected to the decoder.

[0022] S200, obtaining a sequence sample set D, and dividing the sequence sample set into a training set and a test set.

[0023] In the embodiment of the present invention, D may be a time sequence within a set time period. The length of the set time period may be set based on actual needs. Specifically, D={D1, D2, ..., D i ,……,D n}; D i is the i-th time series in D, specifically the time series formed by the occurrence time of the key operation nodes of the flight ground support of the corresponding flight. The value of i ranges from 1 to n, and n is the number of time series; D i=(D i1 , D i2 , ..., D ij , ..., D im ), D ij D i The occurrence time of the jth flight ground support key operation node in , where j ranges from 1 to m, and m is the number of flight ground support key operation nodes.

[0024] In the embodiment of the present invention, the occurrence time of the key operation node of the ground support of the flight is the execution time of the key operation node of the ground support of the flight, which may include the actual block time of the flight, the actual docking time of the flight, the actual cabin door opening time of the flight, the actual start time of unloading, the actual end time of unloading, the actual end time of the refueling truck, the actual start time of the meal, the actual end time of the meal, the actual start time of the baggage loading, the actual end time of the baggage loading, the actual cabin door closing time of the flight, the actual bridge removal time of the flight, and the actual block removal time of the flight. In the embodiment of the present invention, the format of the occurrence time may include year, month, day, hour, minute, and second information, such as 2023-10-21-9:33:00, etc.

[0025] S300: Train the initial flight off-block time prediction model using the training set to obtain a trained flight off-block time prediction model.

[0026] S400, using the test set to test the trained flight off-block time prediction model.

[0027] A method for constructing a flight off-block time prediction model provided by an embodiment of the present invention, wherein the constructed model integrates a frequency-enhanced channel attention module based on discrete cosine transform, a temporal convolutional network module, and a window mixed sparse architecture. The model uses a frequency-enhanced channel attention mechanism to solve the actual flight ground support process and uses Fourier transform to process the Gibbs phenomenon that may appear in the time series, and uses the dilated convolution technology in the temporal convolutional network to extract the features of adjacent time segments of the test point, and integrates the time segment features into the self-attention mechanism, which can capture the dependencies between long-distance time steps, and then put the processed time features into the window mixed sparse architecture. The window mixed sparse architecture can extract and calculate the dominant time segment features while capturing long-term dependencies, thereby reducing the computational complexity and improving the prediction effect.

[0028] Further, in an embodiment of the present invention, the structures and working principles of the first discrete cosine transform-based frequency enhancement channel attention module 11 and the second discrete cosine transform-based frequency enhancement channel attention module 21 are the same, aiming to improve the model's ability to extract frequency features while solving the Gibbs phenomenon and high-frequency noise problems caused by the traditional Fourier transform, and can be the structures and working principles of the existing discrete cosine transform-based frequency enhancement channel attention modules. The working principle may include, for example: performing channel division on the input multivariate time series data, splitting the data into multiple channel variables; performing discrete cosine transform on each channel variable to obtain the corresponding frequency channel vector; stacking and rearranging all frequency channel vectors along the channel dimension to form a frequency channel attention tensor; using a fully connected neural network to learn the frequency channel tensor to obtain a frequency channel attention tensor; performing element-by-element multiplication of the original time series data and the frequency channel attention tensor to obtain a weighted representation; using a projection layer to further process the weighted representation, and controlling the output length for prediction.

[0029] Those skilled in the art know that the specific implementation of the working principle of the frequency-enhanced channel attention module based on discrete cosine transform may be prior art.

[0030] When flight ground support data is interfered by noise caused by various factors at the airport, its waveform may become messy in the time domain and difficult to distinguish from the noise, but it can be relatively easily distinguished in the frequency domain. The present invention processes frequency information through discrete cosine transform, which can essentially avoid the Gibbs phenomenon and inverse transform operations.

[0031] Furthermore, in an embodiment of the present invention, a temporal convolutional network (TCN) module is a deep learning model specifically used to process sequence data. It combines the parallel processing capability of convolutional neural networks and the long-term dependency modeling capability of recurrent neural networks, becoming a powerful tool in sequence modeling tasks. The temporal convolutional network module may include: a one-dimensional causal convolution layer, layer normalization, a Relu activation function, and a dropout layer. The temporal convolutional network uses causal convolution, which is a unidirectional structure. Causal convolution ensures that the current output time step t is only related to the input of time step t and before. This can avoid using future information when predicting the future, not referencing future data, and is more in line with the causal relationship of the time series. The dilated convolution of TCN introduces holes in the convolution kernel, allowing the network to expand the receptive field without increasing the number of parameters, thereby enhancing the model's time perception ability.

[0032] Those skilled in the art know that any method of using a temporal convolutional network module to perform convolution processing on received data to obtain corresponding convolution features falls within the scope of protection of the present invention.

[0033] The accuracy of flight off-block time prediction is crucial. The time of each operation during the support process will fluctuate to varying degrees over time. Therefore, the existence of external factors such as sudden weather changes and peak periods and internal factors such as resource allocation errors and special equipment damage will affect the flight ground support time series. Whether the current support situation is normal depends largely on the current environment. Under this problem, the introduction of TCN can expand the receptive field of the model by adding layers, thereby extracting richer time segment features. In addition, the introduction of TCN is also to make up for the essential defects of the Transformer-like model. The dot product of the query vector and the value vector in the Transformer is a correlation calculation of the data at the time point scale, ignoring the time segment information around the observation point. After integrating TCN, the connection structure of the self-attention mechanism will also be changed. At this time, the input of the query vector and the key vector is the time series data that includes the characteristics of different time segments after TCN processing, while the input of the value vector remains unchanged, and is still the original time series data that has not been processed by TCN.

[0034] Furthermore, if Figure 3 As shown, the encoder 13 includes k encoding modules connected in sequence, wherein each encoding module includes a first window mixed sparse multi-head self-attention layer 1301, a first splicing processing module 1302, a first feedforward neural network layer 1303 and a second splicing processing module 1304 connected in sequence. The decoder 23 includes k decoding modules connected in sequence, each decoding module includes a masked window mixed sparse multi-head self-attention layer 2301, a third splicing processing module 2302, a second window mixed sparse multi-head self-attention layer 2304, a fourth splicing processing module 2305, a second feedforward neural network layer 2306 and a fifth splicing processing module 2307 connected in sequence. It should be noted that Figure 3 Only one encoding module and one decoding module are shown schematically.

[0035] In the embodiment of the present invention, k is a number greater than 1, and k can be set based on actual needs. In an exemplary embodiment, k=6.

[0036] Among them, the first frequency-enhanced channel attention module 11 based on discrete cosine transform and the first time convolutional network module 12 are respectively connected to the window mixed sparse multi-head self-attention layer of the first encoding module, the first time convolutional network module and the first window mixed sparse multi-head self-attention layer of the first encoding module are also respectively connected to the first splicing processing module of the first encoding module, the first splicing processing module of each encoding module is also connected to the corresponding second splicing processing module, and the second splicing processing module of the previous encoding module of two adjacent encoding modules is connected to the first window mixed sparse multi-head self-attention layer and the second splicing processing module of the latter encoding module.

[0037] Among them, the second temporal convolutional network module is respectively connected to the masked window mixed sparse multi-head self-attention layer and the third splicing processing module of the first decoding module, the third splicing processing module of each decoding module is also connected to the corresponding fourth splicing processing module, and the fourth splicing processing module of each decoding module is also connected to the corresponding fifth splicing processing module; the fifth splicing processing module of the first decoding module in two adjacent encoding modules is respectively connected to the masked window mixed sparse multi-head self-attention layer and the third splicing processing module of the second decoding module, and the output of the kth encoding module is respectively used as the input of the second window mixed sparse multi-head self-attention layer of each decoding module.

[0038] Specifically, the output of the first time convolutional network is used as the input of the query vector and key vector of the first window mixed sparse multi-head self-attention layer of the first encoding module, and the output of the first discrete cosine transform-based frequency enhanced channel attention module is used as the input of the value vector of the first window mixed sparse multi-head self-attention layer of the first encoding module; the output of the second time convolutional network is used as the input of the query vector and key vector of the masked window mixed sparse multi-head self-attention layer of the first decoding module, and the output of the second discrete cosine transform-based frequency enhanced channel attention module is used as the input of the value vector of the masked window mixed sparse multi-head self-attention layer of the first decoding module. In the embodiment of the present invention, after the flight ground support data is processed by TCN, the result is used to generate query vectors and key vectors containing the features of different time segments, and the sequence data that has not been processed by TCN is still used as the input of the value vector. When the query vector and key vector generated in this way are calculated for correlation, they can fully associate the effective features between different time segments, mine and utilize the features of important historical data segments, and enhance the modeling effect of time series data.

[0039] Furthermore, in an embodiment of the present invention, the window mixed sparse multi-head self-attention layer is different from the multi-head self-attention layer in the original Transformer in that it includes two parts: local window attention and specific sparse attention. The implementation process of these two parts includes embedding the input sequence features into a high-dimensional space to obtain a query vector Q, a key vector K, a value vector V, scaling the dot product attention to obtain the attention output, and calculating multiple attention heads in parallel and finally merging the output. The difference is that when calculating the dot product of the query vector and the key vector, the correlation between all time points is no longer calculated, but relatively important time points are purposefully selected for calculation. The multi-head self-attention layer in the original Transformer model calculates the correlation between all sequences. Since the correlation between most sequences is low, it contains a large number of invalid calculations. When the sequence length is too long, there will be problems of high computational complexity and low computational efficiency.

[0040] In the embodiment of the present invention, the selection of the dot product calculation of the query vector and the key vector in the window mixed sparse multi-head self-attention layer is composed of two parts: local window attention and specific sparse attention. The local window attention is to screen the observation points near the target point, because in practical applications, the observation points closer to the prediction point have a greater impact on the result and are more important, while the observation points far away from the prediction point contribute less; the specific sparse attention is to screen the time series with higher importance, and select the time series that have a major influence on the prediction point. The window mixed sparse multi-head self-attention layer can reduce the computational complexity, improve the computational efficiency and ensure the computational accuracy.

[0041] In an embodiment of the present invention, the choice of dot product calculation of the query vector and the key vector in the window mixed sparse multi-head self-attention layer in the encoder and the decoder, that is, the first window mixed sparse multi-head self-attention layer, the second window mixed sparse multi-head self-attention layer, and the masked window mixed sparse multi-head self-attention layer can specifically include two parts: local window attention and specific sparse attention.

[0042] Specifically, in an embodiment of the present invention, when the local window attention selects an observation point closer to the prediction point to perform the dot product calculation of the query vector and the key vector, for example, the window size can be set to b, that is, the attention calculation is performed only for the observation points within the window size b, and the data outside the window is filled with a zero vector. The attention matrix obtained in this way is that the dot product calculation of the query vector and the key vector is performed only in the diagonal and the b windows near it (windows of 0.5×b on the left and right sides). For the k-window mixed sparse multi-head self-attention layer, b is a fixed value, each layer is b, the receptive field in the first layer is b, and the value of the observation point in each window in the second layer contains the local information obtained from the b observation points near the observation point at the same position in the first layer. At this time, the receptive field of the second layer is 2b. Therefore, at the kth layer, the size of the receptive field is k×b. Although the bottom layer can obtain little information, the receptive field will become larger and larger as the number of layers increases.

[0043] In an embodiment of the present invention, the specific sparse attention sorts the query vectors by importance, selects the query vectors that contribute more to the calculation and performs dot product calculation with the key vector, thereby ignoring those query vectors with less influence. The local self-attention matrix generated in this way can reduce redundant calculations and improve the efficiency and effect of the model. First, the amount of calculation is reduced by randomly sampling a certain number of key vectors from the key space, and then the correlation between the query vector and these key vectors is calculated. Finally, by selecting multiple key vectors that are most relevant to the query vector, the correlation matrix of these key vectors and the query vector is further calculated. This can effectively improve the calculation efficiency while retaining the information of the key vector.

[0044] In an embodiment of the present invention, the two parts of observation points selected by the window mixed sparse multi-head self-attention layer are spliced ​​and superimposed, and the local window attention is calculated first, and then the specific sparse attention is calculated for the non-overlapping part of the splicing, and finally the calculation result of the overall window mixed sparse multi-head self-attention is obtained.

[0045] Specifically, in an embodiment of the present invention, the first window mixed sparse multi-head self-attention layer, the second window mixed sparse multi-head self-attention layer and the masked window mixed sparse multi-head self-attention layer are used to obtain the attention weight matrix of the time series features, wherein the attention heads therein can be specifically used to perform the following operations: S10, based on the received H time series features, respectively obtain the query vector, key vector and value vector of each time series feature, thereby obtaining H query vectors, H key vectors and H value vectors.

[0046] Those skilled in the art know that the query vector, key vector and value vector of the time series feature may be obtained by the prior art.

[0047] S20, for the rth time series feature, obtain the first attention score vector A1 corresponding to the rth time series feature r =(A1 r1 , A1 r2 , ..., A1 rs , ..., A1 rH ,); get the first attention score matrix A1 corresponding to H time series features, where the size of A1 is H×H, and the value of the r-th row of data in A1 is the first attention score vector of the r-th time series feature; where A1 rs is the attention score of the query vector of the rth time series feature relative to the key vector of the sth time series feature, r ranges from 1 to H, and s ranges from 1 to H; where A1 r Except for the middle set P r Except for the corresponding attention score, the remaining attention scores are 0; r ={A1 r（r-0.5w） , A1 r（r-0.5w+1） , ..., A1 r（r-1） , A1 rr , A1 r（r+1） , ..., A1 r（r+0.5w-1） , A1 r（r+0.5w）}, w is a set value, which can be an odd number greater than 1, and can be an empirical value.

[0048] In an embodiment of the present invention, the attention score of the rth time series feature relative to the sth time series feature may be the dot product between the query vector of the rth time series feature and the key vector of the sth time series feature.

[0049] S30, randomly select Q key vectors from the H key vectors as reference key vectors, and calculate the correlation coefficients between the query vector of the rth time series feature and the Q key vectors to obtain Q correlation coefficients corresponding to the query vector of the rth time series feature.

[0050] In the embodiment of the present invention, Q can be set based on actual needs. In an exemplary embodiment, Q=H / 4.

[0051] In the embodiment of the present invention, the correlation coefficient may be KL divergence.

[0052] S40, taking the maximum correlation coefficient among the Q correlation coefficients corresponding to the query vector of the r-th time series feature as the target correlation coefficient of the query vector of the r-th time series feature; obtaining the target correlation coefficients corresponding to H query vectors; and executing S50.

[0053] S50, sorting the H target correlation coefficients in descending order to obtain sorted H target correlation coefficients, and taking the Z query vectors corresponding to the first Z correlation coefficients of the sorted H target correlation coefficients as Z reference query vectors; and executing S60. Z>1, Z can be an empirical value.

[0054] S60, obtain the second attention score matrix A2 corresponding to the H time series features = (A21, A22, ..., A2 r , ..., A2 H ), where A2 r is the second attention score vector of the rth time series feature, A2 r =(A2 r1 , A2 r2 , ..., A2 rs , ..., A2 rH ), A2 rs is the second attention score of the rth time series feature relative to the sth time series feature, where if the rth time series feature is a time series feature among the Z time series features corresponding to the Z reference query vectors, then A2 rs is the dot product between the query vector of the rth time series feature and the key vector of the sth time series feature. Otherwise, A2 rs =0; execute S70.

[0055] S70, if the rth time series feature is a time series feature among the Z time series features corresponding to the Z reference query vectors, for A1 r Execute the following processing: Traverse A1 r , for the traversed A1 rs , if A1 rs =0, set A1 rs =A2 rs ; Get the processed A1 r , as the final attention score vector of the r-th time series feature; if the r-th time series feature is not a time series feature among the Z time series features corresponding to the Z reference query vectors, A1 r As the final attention score vector of the r-th time series feature.

[0056] Those skilled in the art know that after obtaining the attention score of the time series feature, the attention score can be normalized using the Softmax function and then multiplied with the corresponding value vector to obtain the corresponding attention weight matrix.

[0057] Furthermore, the first to fifth splicing processing modules are used to: splice the received input features to obtain corresponding splicing results, and normalize the splicing results. That is, each splicing module is used for residual connection and layer normalization. Residual connection ensures that information can be directly propagated from forward to backward, avoiding the problem of gradient vanishing in deep networks, and enhancing information transfer, that is, adding the output of the sub-layer (such as self-attention or FFN) to the input to ensure that information is directly transmitted from the input layer to a deeper level, which partially solves the problem of gradient vanishing caused by too many neural network layers. Layer normalization is used to standardize the input matrix, which can accelerate training and improve the stability of the model.

[0058] Those skilled in the art know that the specific structure and working principle of the encoder and decoder may be prior art.

[0059] In an embodiment of the present invention, an encoder is used to extract global time series features, and a decoder is used to combine historical and current information to generate prediction results, thereby effectively utilizing historical data; and the window mixed sparse multi-head self-attention layer can dynamically focus on important historical data points in the time series according to different attention weights, and the multi-head mechanism allows different heads to focus on different time periods or frequency characteristics, thereby having stronger adaptability to time series with greater volatility.

[0060] Furthermore, S300 may specifically include: S301, input the prediction data set in the current batch of sequence samples as the current first training sample data into the first feature encoding module to obtain the corresponding first encoding feature, and input the first encoding feature into the first frequency enhanced channel attention module based on discrete cosine transform; input the to-be-predicted data set in the current batch of sequence samples as the current second training sample data into the second feature encoding module to obtain the corresponding second encoding feature, and input the second encoding feature into the second frequency enhanced channel attention module based on discrete cosine transform; wherein, the first encoding feature includes the data encoding feature, position encoding feature and time encoding feature of the current first training sample data, and the second encoding feature includes the data encoding feature, position encoding feature and time encoding feature of the current second training sample data.

[0061] In an embodiment of the present invention, each batch of sequence samples may include data within a specified time period, such as data within 7 hours. The prediction data set may be sample data within a first set time period, such as data within 6 hours, and the prediction data set is sample data within a second set time period after the prediction data set, such as sample data within the next 1 hour. That is, in an exemplary embodiment, the present invention can use the time series data of the first 6 hours to predict the time series data of the next 1 hour.

[0062] In the embodiment of the present invention, the first coding feature or the second coding feature satisfies the following conditions: FE = α × F + E + P; Among them, FE is the first coding feature or the second coding feature, α is the balance factor, F is the data coding feature, E is the time coding feature, and P is the position coding feature.

[0063] In an embodiment of the present invention, the input training sample data may be a matrix, the size of the matrix is ​​q×m, and q is the number of training sample data. The data of the matrix is ​​preprocessed data. The preprocessing method may be: data=hour+min / 60+sec / 3600, where data is the preprocessed data, hour, min and sec are the hour information, minute information and second information of the occurrence time, for example, the occurrence time is 9:33:40, and the corresponding hour is 9, min is 33, and sec is 40.

[0064] In the embodiment of the present invention, the occurrence time can be feature encoded by the existing data encoding method. For example, the flight ground support time series data is embedded through a one-dimensional convolution operation, the data dimension is mapped, and the input multi-dimensional data is mapped to a certain high-dimensional feature space. In this way, the time series can be mapped into a new feature matrix.

[0065] In the embodiment of the present invention, the position coding feature can satisfy the following conditions: for the k1th odd-numbered bit of data in the rth row vector in the matrix, the coding result is sin(r / 10000) （2k1 / （m-1）） ), for the k2th even-numbered bit of the rth row vector in the matrix, the encoding result is cos(r / 10000) （2k2 / （m-1）） ). The value of r ranges from 1 to q.

[0066] Furthermore, in the embodiment of the present invention, E=E t +E h +E d +E m , E t is the coding feature of the sub-information corresponding to the occurrence time of the key operation node of the ground support of the first flight, E h is the coding feature of the hourly information corresponding to the occurrence time of the key operation node of the first flight ground support, E d is the coding feature of the daily information corresponding to the occurrence time of the key operation node of the first flight ground support, E m The encoding feature of the month information corresponding to the occurrence time of the key operation node of the ground support of the first flight. The month information is used to represent the position of the time window in a year, such as January, February, etc. The day information is used to represent the position of the time window in a month, such as the 2nd, etc.

[0067] Specifically, E t 、E h 、E d and E m You can obtain it by following the steps below: The minute information, hour information, day information and month information of the occurrence time of the key operation node of the ground support of the first flight are normalized respectively to obtain the normalized information.

[0068] Specifically, t =((sec-1) / 59)-0.5, d h =((hour-1) / 23)-0.5, d d =((day-1) / 30)-0.5,d m =((month-1) / 11)-0.5. Among them, d t d h d d and d m The normalized minute information, hour information, day information and month information are respectively represented, day is the day information, and month is the month information. In this way, the minute information, hour information, day information and month information can be converted into values ​​between -0.5 and 0.5.

[0069] The normalized information is mapped to feature dimensions to obtain the corresponding encoding features, that is, E t 、E h 、E d and E m .

[0070] In the embodiment of the present invention, the feature dimension may be 512 dimensions.

[0071] In the embodiment of the present invention, by fusing data coding features, position coding features and time coding features, the extracted features can be made more accurate.

[0072] S302, using a first discrete cosine transform-based frequency enhanced channel attention module to perform feature extraction on the received first coding feature, obtain a corresponding extraction result as the first extracted feature, and input it into the first time convolutional network module and the encoder respectively; and using a second discrete cosine transform-based frequency enhanced channel attention module to perform feature extraction on the received second coding feature, obtain a corresponding extraction result as the second extracted feature, and input it into the second time convolutional network module and the decoder.

[0073] S303, using the first time convolutional network module to perform convolution processing on the received first extracted feature, obtain the corresponding convolution result as the first convolutional feature, and input it to the encoder; and using the second time convolutional network module to perform convolution processing on the received second extracted feature, obtain the corresponding convolution result as the second convolutional feature, and input it to the decoder.

[0074] S304: Use the encoder to perform feature extraction on the received first extraction feature and the first convolution feature, obtain corresponding extraction results as intermediate features, and input them to the decoder.

[0075] S305, using the decoder to perform feature extraction on the received second extracted features, the second convolutional features and the intermediate features, obtain corresponding extraction results as decoding features, and input them into the output prediction module.

[0076] S306: Use the output prediction module to perform feature extraction on the received decoding features to obtain a corresponding prediction result as the current prediction result.

[0077] S307, based on the current prediction result and the corresponding actual result, the loss function value of the current flight off-block time prediction model is obtained. If the loss function value of the current flight off-block time prediction model meets the preset model training end condition, execute S309, otherwise, execute S308; the initial value of the current flight off-block time prediction model is the initial flight off-block time prediction model.

[0078] In the embodiment of the present invention, the actual result is the data set to be predicted.

[0079] In an embodiment of the present invention, the loss function value can be calculated based on an existing loss function, such as a cross entropy loss function. The preset model training end condition can be set based on actual needs. For example, the loss is less than or equal to a set loss threshold and remains unchanged within a set time period.

[0080] S308, updating the parameters of the current flight off-block time prediction model based on the loss function value of the current flight off-block time prediction model, and taking the sequence samples of the next batch as the sequence samples of the current batch, and executing S301.

[0081] S309: Using the current flight off-block time prediction model as the trained flight off-block time prediction model.

[0082] In summary, the flight off-block time prediction model provided by the embodiment of the present invention utilizes TCN to mine time segment features, and can capture the relationship between these segments to make the prediction result closer to the true value, thereby achieving better prediction effect; the frequency enhancement channel attention module based on discrete cosine transform is used to perform frequency enhancement processing on the input data, and frequency domain analysis and feature interaction are performed on each input channel. The network can learn the influence of different frequencies on each channel, while avoiding the Gibbs phenomenon and fully mining important features from the frequency domain, thereby helping the network to efficiently mine potential frequency domain features; the window mixed sparse multi-head self-attention layer performs correlation calculation by screening out time points with higher importance than the prediction point and the neighboring points of the prediction point, which can reduce the computational complexity, improve the computational efficiency and ensure the computational accuracy.

[0083] Another embodiment of the present invention provides a flight off-block time prediction method, which is implemented based on the flight off-block time prediction model obtained by the method provided in the above embodiment, and the prediction method includes: S1, obtaining a time series data set for prediction, wherein the time series data set for prediction includes the occurrence times of key operation nodes of flight ground support of multiple flights within a specified time period.

[0084] S2, inputting the prediction time series data set into the flight off-block time prediction model to obtain a corresponding prediction result, and then obtaining a predicted value of the flight off-block time that needs to be predicted.

[0085] 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.

[0086] 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.

[0087] 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.

[0088] 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 improvement 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 off-block time prediction model, characterized in that: The method comprises the following steps: S100, constructing an initial flight off-block time prediction model; the model includes: a first processing model and a second processing model, the first processing model includes a first feature encoding module, a first discrete cosine transform-based frequency enhancement channel attention module, a first time convolutional network module and an encoder connected in sequence, the second processing model includes a second feature encoding module, a second discrete cosine transform-based frequency enhancement channel attention module, a second time convolutional network module, a decoder and an output prediction module connected in sequence, wherein the encoder is connected to the decoder, the first discrete cosine transform-based frequency enhancement channel attention module is also connected to the encoder, and the second discrete cosine transform-based frequency enhancement channel attention module is also connected to the decoder; S200, obtain a sequence sample set D, and divide the sequence sample set into a training set and a test set; D = {D1, D2, ..., D i ,……,D n }; D i is the i-th time series in D, i ranges from 1 to n, and n is the length of the time series; D i =(D i1 , D i2 , ..., D ij , ..., D im ), D ij D i The occurrence time of the jth flight ground support key operation node in , where j ranges from 1 to m, and m is the number of flight ground support key operation nodes; S300, training an initial flight off-block time prediction model using a training set to obtain a trained flight off-block time prediction model; S400, using the test set to test the trained flight off-block time prediction model.

2. The method according to claim 1, characterized in that: S300 specifically includes: S301, input the prediction data set in the current batch of sequence samples as the current first training sample data into the first feature encoding module to obtain the corresponding first encoding feature, and input the first encoding feature into the first frequency enhancement channel attention module based on discrete cosine transform; input the to-be-predicted data set in the current batch of sequence samples as the current second training sample data into the second feature encoding module to obtain the corresponding second encoding feature, and input the second encoding feature into the second frequency enhancement channel attention module based on discrete cosine transform; wherein the first encoding feature includes the data encoding feature, position encoding feature and time encoding feature of the current first training sample data, and the second encoding feature includes the data encoding feature, position encoding feature and time encoding feature of the current second training sample data; S302, using a first discrete cosine transform-based frequency enhanced channel attention module to perform feature extraction on the received first coding feature, obtain a corresponding extraction result as a first extracted feature, and input it into the first time convolutional network module and the encoder respectively; and using a second discrete cosine transform-based frequency enhanced channel attention module to perform feature extraction on the received second coding feature, obtain a corresponding extraction result as a second extracted feature, and input it into the second time convolutional network module and the decoder; S303, using the first time convolutional network module to perform convolution processing on the received first extracted feature, obtain a corresponding convolution result as the first convolutional feature, and input it to the encoder; and using the second time convolutional network module to perform convolution processing on the received second extracted feature, obtain a corresponding convolution result as the second convolutional feature, and input it to the decoder; S304, using the encoder to perform feature extraction on the received first extraction feature and the first convolution feature, obtain corresponding extraction results as intermediate features, and input them to the decoder; S305, using the decoder to perform feature extraction on the received second extracted features, the second convolutional features and the intermediate features, obtain corresponding extraction results as decoding features, and input them into the output prediction module; S306, using the output prediction module to perform feature extraction on the received decoding features to obtain a corresponding prediction result as the current prediction result; S307, based on the current prediction result and the corresponding actual result, the loss function value of the current flight off-block time prediction model is obtained. If the loss function value of the current flight off-block time prediction model meets the preset model training end condition, S309 is executed; otherwise, S308 is executed; the initial value of the current flight off-block time prediction model is the initial flight off-block time prediction model; S308, updating the parameters of the current flight off-block time prediction model based on the loss function value of the current flight off-block time prediction model, and taking the sequence samples of the next batch as the sequence samples of the current batch, and executing S301; S309: Using the current flight off-block time prediction model as the trained flight off-block time prediction model.

3. The method according to claim 2, characterized in that The encoder includes k encoding modules connected in sequence, wherein each encoding module includes a first window mixed sparse multi-head self-attention layer, a first splicing processing module, a first feedforward neural network layer and a second splicing processing module connected in sequence, wherein the first discrete cosine transform-based frequency enhanced channel attention module and the first time convolution network module are respectively connected to the window mixed sparse multi-head self-attention layer of the first encoding module, the first time convolution network module and the first window mixed sparse multi-head self-attention layer of the first encoding module are also respectively connected to the first splicing processing module of the first encoding module, the first splicing processing module of each encoding module is also connected to the corresponding second splicing processing module, and the second splicing processing module of the previous encoding module of two adjacent encoding modules is respectively connected to the first window mixed sparse multi-head self-attention layer and the first splicing processing module of the next encoding module; k>1; The decoder includes k decoding modules connected in sequence, each decoding module includes a masked window mixed sparse multi-head self-attention layer, a third splicing processing module, a second window mixed sparse multi-head self-attention layer, a fourth splicing processing module, a second feedforward neural network layer and a fifth splicing processing module connected in sequence, wherein the second temporal convolutional network module is respectively connected to the masked window mixed sparse multi-head self-attention layer and the third splicing processing module of the first decoding module, the third splicing processing module of each decoding module is also connected to the corresponding fourth splicing processing module, and the fourth splicing processing module of each decoding module is also connected to the corresponding fifth splicing processing module; the fifth splicing processing module of the first decoding module of two adjacent decoding modules is respectively connected to the masked window mixed sparse multi-head self-attention layer and the third splicing processing module of the second decoding module, and the output of the kth encoding module is respectively used as the input of the second window mixed sparse multi-head self-attention layer of each decoding module; Among them, the first splicing processing module to the fifth splicing processing module are all used to: splice the received input features to obtain corresponding splicing results, and normalize the splicing results.

4. The method according to claim 3, characterized in that The output of the first temporal convolutional network is used as the input of the query vector and key vector of the first window mixed sparse multi-head self-attention layer of the first encoding module, and the output of the first discrete cosine transform-based frequency enhanced channel attention module is used as the input of the value vector of the first window mixed sparse multi-head self-attention layer of the first encoding module; the output of the second temporal convolutional network is used as the input of the query vector and key vector of the masked window mixed sparse multi-head self-attention layer of the first decoding module, and the output of the second discrete cosine transform-based frequency enhanced channel attention module is used as the input of the value vector of the masked window mixed sparse multi-head self-attention layer of the first decoding module.

5. The method according to claim 4, characterized in that The attention heads in the first window mixed sparse multi-head self-attention layer, the second window mixed sparse multi-head self-attention layer, and the masked window mixed sparse multi-head self-attention layer are specifically used to perform the following operations: S10, based on the received H time series features, respectively obtain the query vector, key vector and value vector of each time series feature; obtain H query vectors, H key vectors and H value vectors; S20, obtain the first attention score matrix A1 corresponding to H time series features = (A11, A12, ..., A1 r , ..., A1 H ); Among them, A1 r is the first attention score vector of the rth time series feature among the H time series features, r ranges from 1 to H, and the initial value is 1; where A1 r =(A1 r1 , A1 r2 , ..., A1 rs , ..., A1 rH ), A1 rs is the first attention score of the rth time series feature relative to the sth time series feature, s ranges from 1 to H, and the initial value is 1; where A1 r Except for the middle set P r Except for the corresponding attention score, the remaining attention scores are 0; r ={A1 r（r-0.5w） , A1 r（r-0.5w+1） , ..., A1 r（r-1） , A1 rr , A1 r（r+1） , ..., A1 r（r+0.5w-1） , A1 r（r+0.5w） }, w is the set value; S30, randomly selecting Q key vectors from the H key vectors as reference key vectors, and calculating the correlation coefficients between the query vector of the r-th time series feature and the Q key vectors, to obtain Q correlation coefficients corresponding to the query vector of the r-th time series feature; S40, taking the maximum correlation coefficient among the Q correlation coefficients corresponding to the query vector of the r-th time series feature as the target correlation coefficient of the query vector of the r-th time series feature; obtaining the target correlation coefficients corresponding to H query vectors; executing S50; S50, sorting the H target correlation coefficients in descending order to obtain sorted H target correlation coefficients, and using Z query vectors corresponding to the first Z correlation coefficients of the sorted H target correlation coefficients as Z reference query vectors; executing S60; S60, obtain the second attention score matrix A2 corresponding to the H time series features = (A21, A22, ..., A2 r , ..., A2 H ), where A2 r is the second attention score vector of the rth time series feature, A2 r =(A2 r1 , A2 r2 , ..., A2 rs , ..., A2 rH ), A2 rs is the second attention score of the rth time series feature relative to the sth time series feature, where if the rth time series feature is a time series feature among the Z time series features corresponding to the Z reference query vectors, then A2 rs is the dot product between the query vector of the rth time series feature and the key vector of the sth time series feature. Otherwise, A2 rs =0; execute S70; S70, if the rth time series feature is a time series feature among the Z time series features corresponding to the Z reference query vectors, for A1 r Execute the following processing: Traverse A1 r , for the traversed A1 rs , if A1 rs =0, set A1 rs =A2 rs ; Get the processed A1 r , as the final attention score vector of the r-th time series feature; if the r-th time series feature is not a time series feature among the Z time series features corresponding to the Z reference query vectors, A1 r As the final attention score vector of the r-th time series feature.

6. The method according to claim 2, characterized in that The first coding feature or the second coding feature meets the following conditions: FE = α × F + E + P; Among them, FE is the first encoding feature or the second encoding feature, α is the balance factor, F is the data encoding feature, E is the time encoding feature, and P is the position encoding feature; Where E = E t +E h +E d +E m , E t is the coding feature of the sub-information corresponding to the occurrence time of the key operation node of the ground support of the first flight, E h is the coding feature of the hourly information corresponding to the occurrence time of the key operation node of the first flight ground support, E d is the coding feature of the daily information corresponding to the occurrence time of the key operation node of the first flight ground support, E m It is the coding feature of the monthly information corresponding to the occurrence time of the key operation node of the ground support of the first flight.

7. The method according to claim 1, characterized in that The occurrence time of the key operation nodes of the flight ground support includes: the actual wheel block time of the flight, the actual docking time of the flight, the actual cabin door opening time of the flight, the actual start time of unloading, the actual end time of unloading, the actual end time of the refueling truck, the actual start time of meals, the actual end time of meals, the actual start time of baggage loading, the actual end time of baggage loading, the actual cabin door closing time of the flight, the actual bridge removal time of the flight and the actual wheel block removal time of the flight.

8. An electronic device, characterized in that: including a processor and a memory; The processor is used to execute the steps of the method according to any one of claims 1 to 7 by calling the program or instruction stored in the memory.

9. A computer-readable storage medium, characterized in that: The computer-readable storage medium is used to store a program or an instruction, wherein the program or the instruction enables a computer to execute the steps of the method according to any one of claims 1 to 7.

Images