CN116910919A - Filling method and device under Gao Queshi rate of aircraft track - Google Patents
Filling method and device under Gao Queshi rate of aircraft track Download PDFInfo
- Publication number
- CN116910919A CN116910919A CN202311154958.1A CN202311154958A CN116910919A CN 116910919 A CN116910919 A CN 116910919A CN 202311154958 A CN202311154958 A CN 202311154958A CN 116910919 A CN116910919 A CN 116910919A
- Authority
- CN
- China
- Prior art keywords
- track
- missing
- dimensional
- filling
- data
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
- 238000000034 method Methods 0.000 title claims abstract description 38
- 238000012549 training Methods 0.000 claims abstract description 26
- 238000012360 testing method Methods 0.000 claims abstract description 11
- 238000012795 verification Methods 0.000 claims abstract description 10
- 230000006870 function Effects 0.000 claims description 14
- 239000011159 matrix material Substances 0.000 claims description 14
- 238000013528 artificial neural network Methods 0.000 claims description 7
- 238000004364 calculation method Methods 0.000 claims description 5
- 230000015654 memory Effects 0.000 claims description 5
- 238000012545 processing Methods 0.000 claims description 5
- 230000007246 mechanism Effects 0.000 claims description 4
- 238000010606 normalization Methods 0.000 claims description 4
- ORILYTVJVMAKLC-UHFFFAOYSA-N Adamantane Natural products C1C(C2)CC3CC1CC2C3 ORILYTVJVMAKLC-UHFFFAOYSA-N 0.000 claims description 3
- 238000004140 cleaning Methods 0.000 claims description 3
- 230000006798 recombination Effects 0.000 claims description 3
- 238000005215 recombination Methods 0.000 claims description 3
- 230000009467 reduction Effects 0.000 claims description 3
- 230000007547 defect Effects 0.000 abstract description 2
- 238000010586 diagram Methods 0.000 description 16
- 238000012217 deletion Methods 0.000 description 10
- 230000037430 deletion Effects 0.000 description 10
- 238000013461 design Methods 0.000 description 10
- 230000008569 process Effects 0.000 description 7
- 238000013507 mapping Methods 0.000 description 4
- 238000011176 pooling Methods 0.000 description 4
- 238000013527 convolutional neural network Methods 0.000 description 3
- 230000000873 masking effect Effects 0.000 description 3
- 238000004891 communication Methods 0.000 description 2
- 238000013135 deep learning Methods 0.000 description 2
- 238000012806 monitoring device Methods 0.000 description 2
- 230000000306 recurrent effect Effects 0.000 description 2
- 238000005070 sampling Methods 0.000 description 2
- 101100063942 Neurospora crassa (strain ATCC 24698 / 74-OR23-1A / CBS 708.71 / DSM 1257 / FGSC 987) dot-1 gene Proteins 0.000 description 1
- 230000004913 activation Effects 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 230000008859 change Effects 0.000 description 1
- 238000007621 cluster analysis Methods 0.000 description 1
- 238000010276 construction Methods 0.000 description 1
- 230000003247 decreasing effect Effects 0.000 description 1
- 238000013136 deep learning model Methods 0.000 description 1
- 230000000694 effects Effects 0.000 description 1
- 238000000605 extraction Methods 0.000 description 1
- 238000005429 filling process Methods 0.000 description 1
- 230000007257 malfunction Effects 0.000 description 1
- 238000004519 manufacturing process Methods 0.000 description 1
- 239000003550 marker Substances 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 238000012544 monitoring process Methods 0.000 description 1
- 208000033937 musculocontractural type Ehlers-Danlos syndrome Diseases 0.000 description 1
- 210000005036 nerve Anatomy 0.000 description 1
- 210000002569 neuron Anatomy 0.000 description 1
- 230000008520 organization Effects 0.000 description 1
- 230000000644 propagated effect Effects 0.000 description 1
- 238000011084 recovery Methods 0.000 description 1
- 230000006403 short-term memory Effects 0.000 description 1
- 239000007787 solid Substances 0.000 description 1
- 230000009466 transformation Effects 0.000 description 1
- 238000013519 translation Methods 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F30/00—Computer-aided design [CAD]
- G06F30/10—Geometric CAD
- G06F30/15—Vehicle, aircraft or watercraft design
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/22—Matching criteria, e.g. proximity measures
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F30/00—Computer-aided design [CAD]
- G06F30/20—Design optimisation, verification or simulation
- G06F30/27—Design optimisation, verification or simulation using machine learning, e.g. artificial intelligence, neural networks, support vector machines [SVM] or training a model
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02P—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN THE PRODUCTION OR PROCESSING OF GOODS
- Y02P90/00—Enabling technologies with a potential contribution to greenhouse gas [GHG] emissions mitigation
- Y02P90/30—Computing systems specially adapted for manufacturing
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Evolutionary Computation (AREA)
- General Engineering & Computer Science (AREA)
- Geometry (AREA)
- Data Mining & Analysis (AREA)
- Artificial Intelligence (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Computer Hardware Design (AREA)
- Evolutionary Biology (AREA)
- Bioinformatics & Computational Biology (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Aviation & Aerospace Engineering (AREA)
- Medical Informatics (AREA)
- Software Systems (AREA)
- Mathematical Analysis (AREA)
- Computational Mathematics (AREA)
- Automation & Control Theory (AREA)
- Pure & Applied Mathematics (AREA)
- Mathematical Optimization (AREA)
- Image Processing (AREA)
Abstract
The invention discloses a filling method and a filling device under Gao Queshi rate of an aircraft track, wherein the method comprises the following steps: s1, extracting track data comprising time, longitude, latitude and altitude; s2, normalizing the extracted track data; s3, matching the normalized track data with the missing modes, selecting the track data in each missing mode as a track data set according to the same proportion after matching is completed, and dividing the track data set into a training set, a test set and a verification set according to the proportion; s4, inputting a missing track into a track missing filling model, outputting filled track data by the model, and determining the model by the following steps: and constructing a track missing filling model comprising a track embedding module with missing marks, a lossless feature encoding module and an asymmetric track feature decoding module, and training the track missing filling model according to a training set, a testing set and a verification set. The defect that the track filling precision is low and the filling speed is low for high missing rate is overcome.
Description
Technical Field
The invention relates to the field of filling of aircraft track data loss, in particular to a filling method and device under Gao Queshi rate of an aircraft track.
Background
Due to system errors or other malfunctions, the flight positions collected by the monitoring devices are typically lost or invalidated at some different time or for successive periods of time. While the popularity of ADS-B devices has greatly reduced the difficulty of target monitoring in air traffic, it has not been widely used in many developing countries due to economic or safety concerns. Furthermore, the reliability of communication of ADS-B devices has also been one of the major obstacles to their use. In practice, since the lack of aircraft tracks is generally continuous, recovery of the monitoring devices in ATCSs is very difficult, which presents a significant challenge to existing interpolation methods.
The deep learning-based track filling method is to design a deep neural network structure, and extract high-level abstract features through the combination of bottom-layer network neurons. Most of the methods also adopt a convolutional neural network as a backbone network, so that the scale of model parameters is greatly reduced by utilizing the inherent local connectivity and weight sharing of the convolutional neural network. However, the problems of incomplete characteristic experience of the convolutional neural network and characteristic loss caused by the adoption of a pooling layer are difficult to avoid. Another type of model class generating method is to set a generator for generating the filling track and a discriminator for judging the random input of the filling track and the real track. Most depth generation models of such methods are currently autoregressive, i.e. they model the value of the current time step from previous tracks. However, under the condition of high loss rate of tracks, modeling needs to be performed on a long-time track sequence, accumulated errors usually occur in the model, the accuracy of filled track points is low, and practical requirements are difficult to meet. Obviously, the prior track filling method based on deep learning faces the bottleneck of insufficient performance due to the feature representation capability, and cannot cope with increasingly complex airspace environment and flight operation modes of aircrafts. Meanwhile, since the conventional autoregressive models, such as a recurrent neural network (Recurrent Neural Networks, RNN) and a Long Short-Term Memory (LSTM), need to sequentially process the input track point data, each input track point must be processed recursively, which results in that the model needs multiple recursion operations to complete the filling of the missing track, resulting in larger calculation cost and lower filling efficiency. Furthermore, the design of these traditional autoregressive class models is often complex, meaning that they have more parameters, further resulting in increased inference time.
Disclosure of Invention
The invention aims to overcome the defects of low track filling precision and low filling speed for high missing rate in the prior art and provides a method and a device for filling a track Gao Queshi rate of an aircraft.
In order to achieve the above object, the present invention provides the following technical solutions:
a method of filling up an aircraft track Gao Queshi rate comprising the steps of:
s1, extracting track data comprising time, longitude, latitude and altitude;
s2, normalizing the extracted track data;
s3, matching the normalized track data with the missing modes, selecting the track data in each missing mode as a track data set according to the same proportion after matching is completed, and dividing the track data set into a training set, a test set and a verification set according to the proportion;
s4, inputting a missing track into the track missing filling model, outputting filled track data by the track missing filling model, and determining the track missing filling model by the following steps: and constructing a track loss filling model comprising a track embedding module with a loss identification, a lossless feature encoding module and an asymmetric track feature decoding module, and training the track loss filling model by using the training set, the testing set and the verification set in the step S3.
Preferably, in step S1, when track data including time, longitude, latitude and altitude is extracted, track message data is first extracted from the ATC system, the track message data is decoded, and the decoded data is subjected to data cleaning.
Preferably, in step S3, the missing modes include a take-off missing mode, a fly-flat missing mode, a landing missing mode, and a random missing mode.
Preferably, in step S4, the track embedding module with the missing mark includes a block embedding component and a missing position encoding component, where the block embedding component receives the missing track sequence X and organizes each known track point therein into a track matrix T according to a time sequence, the track matrix T is tensor flattened to become a column vector V of one dimension and input to the full-connection layer, and the missing track sequence X is converted into a high-dimensional missing track sequenceThe missing position coding component adds different coding information to the known and unknown track points respectively, the missing position coding component comprises a missing mark coder and a fixed position coder, and the missing position coder gives a high-dimensional missing track sequence->The track points in the track are added with one-dimensional missing mark codes, and the codes are described as follows by a formula:
，
wherein ,encoded track representation for adding one-dimensional deletion identity, < >>，For a vector space with one dimension (L, (d+1)), d represents the high-dimensional missing track sequence +.>Dimension of (2)The value L is the high-dimensional deletion track sequence +.>D and L are positive integers, and the fixed position encoder gives the high-dimensional missing track sequence +.>Adding fixed position codes to track points in the track points, and describing the fixed position codes by a formula as follows:
，
wherein ,representing a fixed position coding matrix->Element values on row i, row 2j,
will be and />Adding, described by the formula:
，
wherein P is a high-dimensional track vector added with a position coding component,is->And a fixed position coding matrix with consistent dimension.
Preferably, in step S4, the lossless feature-encoding module comprises N Transformer Block,
n is equal to or greater than 2 and is an integer, and is used for learning the high-dimensional characteristic representation of the track point from the high-dimensional track vector P, and the high-dimensional characteristic representation is described by the formula:
，
wherein E represents a high-dimensional feature sequence learned from the high-dimensional track vector P;
transformer Block includes a multi-head group attention component, a multi-layer sensor, and 2-layer normalization.
Preferably, in step S4, E is added to the fixed position encoded elements and the result is transmitted to an asymmetric track feature decoding module, formulated as:
，
wherein ,representing a high-dimensional feature sequence with E added to the fixed-position coded feature.
Preferably, the asymmetric track feature decoding module and the lossless track feature encoding module are mutually independent, and the internal nerve parameter size and the module size of the asymmetric track feature decoding module are different, the asymmetric track feature decoding module comprises 1 Transformer Block and 1 multi-layer perceptron, transformer Block pairs of the asymmetric track feature decoding moduleDecoding to generate a high-dimensional characteristic representation sequence of a complete track by adopting a global self-attention mechanism; the multi-layer perceptron performs dimension reduction processing on the generated complete track high-dimensional characteristic representation sequence to map to a low-dimensional characteristic vector space and then perform characteristic recombination, and finally outputs the complete track sequence.
Preferably, in step S4, a loss function, a self-supervision training method and an Adam optimizer are used for model training, and the loss function calculates that the track points adopted by the loss are different in different missing modes, and the calculation formula of the loss function is as follows:
，
wherein ,for the true value of the j-th element of the i-th track point,/for the j-th element>The model output value of the j element of the ith track point is represented by longitude, latitude and height, H is the total number of track points taken by the track missing filling model according to a missing mode, the positive integer is represented by M, the number of elements of each track point is represented by constant 3.
An aircraft track loss padding apparatus comprising at least one processor and at least one memory communicatively coupled to the processor, the memory storing instructions for execution by the at least one processor to enable the at least one processor to perform any of the steps of the method.
Compared with the prior art, the invention has the beneficial effects that:
1. the invention fills the missing track points by constructing the track missing filling model comprising the track embedding module with the missing mark, the lossless feature encoding module and the asymmetric track feature decoding module, thereby solving the problems that the filled track points have low precision and do not meet the actual requirements;
2. the missing position coding component in the track embedding module with the missing mark adds different coding information to the known track points and the unknown track points respectively, so that priori knowledge is fully utilized, the unknown track information is recovered as much as possible by the model, and the track filling precision is improved;
3. the lossless feature coding module comprises a multi-head group attention component, a multi-layer perceptron and 2-layer standardization without a pooling layer, so that the problem of track point feature loss caused by the pooling layer is avoided, and the track filling precision is further improved;
4. compared with the lossless feature encoding module, the asymmetric track feature decoding module is asymmetric in design in terms of module size and module parameters, and can well complete the task of reconstructing tracks, meanwhile, the asymmetric track feature decoding module only comprises 1 piece Transformer Block and 1 multi-layer perceptron, so that the asymmetric track feature decoding module has the characteristic of light weight, and further the filling speed is improved.
Drawings
FIG. 1 is a flow chart of the invention;
FIG. 2 is a block diagram of a track filling model;
FIG. 3 is a block diagram of a track embedding module with missing identifications;
FIG. 4 is a block diagram of a lossless feature encoding module;
FIG. 5 is a block diagram of a first portion of an asymmetric track feature decoding module;
FIG. 6 is a block diagram of a second portion of the asymmetric track feature decoding module;
FIG. 7 is a diagram of a random track loss mask pattern;
FIG. 8 is a diagram of a flat flight path miss mask pattern;
FIG. 9 is a diagram of a take-off track loss mask pattern;
FIG. 10 is a pattern diagram of a drop-out mask;
FIG. 11 is a schematic diagram of a random missing track;
FIG. 12 is a simplified flowchart of an example track filling.
Detailed Description
The present invention will be described in further detail with reference to test examples and specific embodiments. It should not be construed that the scope of the above subject matter of the present invention is limited to the following embodiments, and all techniques realized based on the present invention are within the scope of the present invention.
Examples
As shown in fig. 1, a filling method for an aircraft track Gao Queshi rate includes the following steps:
s1, extracting track data comprising time, longitude, latitude and altitude;
s2, normalizing the extracted track data;
s3, matching the normalized track data with the missing modes, selecting the track data in each missing mode as a track data set according to the same proportion after matching is completed, and dividing the track data set into a training set, a test set and a verification set according to the proportion;
s4, inputting a missing track into the track missing filling model, outputting filled track data by the track missing filling model, and determining the track missing filling model by the following steps: and constructing a track loss filling model comprising a track embedding module with a loss identification, a lossless feature encoding module and an asymmetric track feature decoding module, and training the track loss filling model by using the training set, the testing set and the verification set in the step S3.
In step S1, track message data is led from the real ATC system, the track message data is decoded, four elements including time, longitude, latitude and altitude are extracted, and other irrelevant element data is removed, so that the situation that the data is incomplete or even wrong in the information screened from a large number of messages is necessary due to various other external factors or network communication faults or terminal acquisition equipment failure. For this, the decoded track information needs to be further cleaned, and the specific cleaning process is as follows: discarding 95% of the number of midway tracks which are less than the due total sampling number from the take-off time to the landing time in the track data, and interpolating adjacent points of the missing track points to form complete track data for the tracks which meet more than 95%. And removing the missing and invalid data according to the rule jointly met by the four elements of the track, so as to ensure that each piece of track data contains four elements of time, longitude, latitude and altitude and has normal acquisition values.
Track data determination example: and a large amount of track data, wherein each section of track data is divided into A track sections with fixed length. Assuming that a certain flight has a flight duration of 2 hours, wherein 200 track points are required, taking the tracks of 200 track points as a basic data unit, and the data dimension is 1 x 200; each track point includes three-dimensional data representing longitude, latitude, and altitude, respectively, then the individual track segment data units X, the data dimensions being organized as 3X 1X 200. The channel dimension is 3, which represents three values of x, y and z, the dimension of a single channel is 1 x 200, which represents two hundred track points which are arranged in time sequence. When some track points are only some of the data are missing or invalid, the points are deleted as well, and only the track points with valid data are reserved, so that the data Y to be encoded are obtained, wherein the dimension is (3, 1, W), and the value of W is a positive integer less than 200.
In step S2, the specific contents of normalizing the extracted track data are: and sampling the track data once every 4 seconds, and performing interpolation processing on the points with the missing samples by using the front and rear points. And then carrying out element level normalization on the track data after interpolation, and removing data expression inconsistency caused by dimension.
In step S3, the normalized track data in step S2 is divided into a training set, a test set and a verification set required by model training in the proportions of 80%, 10% and 10%, cluster analysis is performed from a large number of missing tracks, four missing modes according to the real track missing condition are summarized, a take-off missing mode, a flat flight missing mode, a landing missing mode and a random missing mode, and the tracks in the complete track data are downsampled according to the four modes, so that the trained data are attached to the real missing condition, and each missing mode is specifically defined as follows: let the number of track points formed by continuously missing track segments in the down-sampled track be c, the number of all track points be A, and the method comprises the following steps: when c is more than 30% of A, the continuous missing track segment height is increased and the initial track point height is less than 8000 m, the missing mode is take-off missing; when c is more than 30% of A, and the heights of the starting track point and the ending track point of the continuous missing track segment are more than 8000 meters, the missing mode is a flat flight missing; when c is more than 30% of A, the height of the continuous missing track segment is decreased and the height of the initial track point is less than 8000 m, the missing mode is landing missing; if any of the above-described deletion patterns is not satisfied, the deletion pattern is random deletion.
In step S4, a track loss filling model including a track embedding module with a loss-free identifier, a lossless feature encoding module and an asymmetric track feature decoding module is constructed, as shown in fig. 2, it can be seen from fig. 2 that the constructed track loss filling model further includes a position encoding (element-wise), and the specific construction process is as follows:
the track embedding module with the missing marks is used for mapping the low-dimensional track features to the high-dimensional vector space with time position information. Firstly, each known track point in a single track is organized into a matrix T according to a time sequence, so that each row represents track information at one moment, and the specific organization form is as follows:
，
wherein ,time value representing the nth moment, +.>、/>、/>Values representing longitude, latitude, and altitude after corresponding normalization at the nth time; the track matrix T is then tensor flattened (flat) such that the track matrix T is flattened to a dimension +.>Is defined by the column vector V; finally, the column vector V is input into a full connection layer (Fully Connected Layer), so that the mapping of the low-dimensional track characteristic to the high-dimensional vector space is realized.
The track Embedding module with the missing mark is composed of two components, namely a block Embedding (Patch Embedding) and a missing position coding (Missing Positional Embedding), wherein the Patch Embedding is used for mapping missing track sequence data to a high-latitude characteristic space, and the missing position coding is used for capturing time sequence relation characteristic information between original missing tracks. After obtaining a high-dimensional representation of the tracks, the missing position-coding component may further enrich the inter-track association information. Position of absenceThe coding component consists of a missing mark code and a fixed position code, and different coding information, such as 0/1 coding, is respectively added to the known and unknown track points, so that priori knowledge is fully utilized, and the model recovers the unknown track information as much as possible. The fixed position coding is to make the model pay attention to the relation between the track time sequences, and the embodiment adopts a fixed position coding implementation method based on sine functions and cosine functions. For a high-dimensional track representation sequence of input length LThe deletion marker codes for the sequence +.>The one dimension is used for representing whether the deletion exists, and the formula is as follows:
，
wherein ,encoded track representation for adding one-dimensional deletion identity, < >>，For a vector space with one dimension (L, (d+1)), d represents the high-dimensional missing track sequence +.>L is the high-dimensional missing track sequence +.>D and L are positive integers,
the fixed position coding formula can be described as:
，
wherein ,representing a fixed position coding matrix->Element values on the ith row, 2j column of the system, and then representing the track encoded by the fixed position encoding matrix and the added deletion mark +.>The corresponding position elements are added, and the formula can be expressed as:
wherein P is a high-dimensional track vector added with a position coding component,is->A fixed position coding matrix with consistent dimension;
as shown in FIG. 3, the structure diagram of the track embedding module with missing marks is shown, and a missing track sequence is given
，
The functionality of the track embedding module with missing identifications can be summarized as:
，
wherein P represents a vector space representation sequence mapping the missing track sequence X from the original data to a high dimension,
the above references to P refer to the same content, but are different in terms of different flow positions, and are substantially identical.
The lossless track feature coding module is used for learning high-dimensional feature representation of track points, constructing a track sequence model, capturing an implicit flight motion mode among the track points, wherein the lossless track feature coding module is formed by linearly stacking a plurality of Transformer Block blocks in series, transformer Block blocks are deep neural network modules based on a self-attention mechanism, and a structural diagram is shown in fig. 4. In order to solve the problem of low filling precision under the condition of large-area missing of the flight path, the lossless flight path feature coding module designed by the embodiment cancels a pooling layer so as to avoid feature loss caused by the characteristic extraction of the missing flight path vector by the coding module; the multi-head self-attention component is used, so that the coding module can capture the inherent complex relation among the track points, and particularly when the track loss rate is high, the coding module can fully extract various relation norms existing among the track points; and the coding module starts from the perspective of the global track points by using the multi-layer perceptron, and acquires the global relation characteristics of the track points and the track point set. The characteristic representation quality of the track directly relates to the subsequent filling effect, so the invention obtains high-quality high-dimensional track characteristic representation by stacking a plurality of transform blocks. For a given high-dimensional track feature representation P, the function of the lossless track feature encoding module can be summarized as:
，
where E represents a high-dimensional feature sequence learned from the track vector space sequence P.
The asymmetric track feature decoding module comprises a Transformer Block block and a Multi-Layer perceptron (MLP) component. As shown in fig. 5 and fig. 6, which are schematic diagrams of an asymmetric track feature decoding module according to an embodiment, fig. 5 and fig. 6 are schematic diagrams of an asymmetric track feature decoding module according to an embodiment, wherein a second layer standardized block in fig. 5 is connected with a hidden layer 1 in fig. 6, and a connection arrow points from the second layer standardized block to the hidden layer 1, and the two blocks are connected, so that the schematic diagrams of the asymmetric track feature decoding module according to the embodiment are completely formed in fig. 5 and fig. 6. Unlike a typical decoder, the asymmetric track feature decoding module of the present invention is a separate and asymmetric design, i.e., completely independent of the encoder and with different magnitudes of the neural network parameters within the module. This is determined by the nature of the track information, which is typically quite low in information density, each track point being quite closely related to the track between adjacent time points, which is less semantically related to the neural network. Second, the decoding module functions to map the potential representation of the track back to the track sequence, which is essentially the sequence of reconstructed track points, resulting in a lower semantic level of the decoding module output. Therefore, the invention designs a decoding module which is asymmetric with the track characteristic encoder in terms of module size and module parameters, not only can better complete the task of reconstructing the track, but also can realize high efficiency due to the light weight design of the decoding module. In general, the asymmetry is illustrative of the decoder and encoder asymmetry, the encoder design is large and full for fully extracting the potential representation of the track features, while the decoder design is very lightweight because the semantic information required for the padding task of decoding is low-level, so the decoder can be designed very simply. The asymmetric design can speed up the model's inference at the completion of the flight path tasks, the linear rectification function (Rectified Linear Unit) in fig. 6 being the activation function.
Before the encoded missing track features are input into the asymmetric track feature decoding module, the elements which are encoded with fixed positions need to be added again in advance, so that the position feature information of the missing part of the track is added, and the feature correlation of the missing part of the track and the known track is enhanced. The calculation process can be summarized as follows:
，
wherein ,representing a high-dimensional feature sequence with the addition of position-coded features to the original feature E. The invention provides an asymmetric track feature decoding module which mainly aims at carrying out feature processing with position informationAnd decoding the rows, reducing the dimension, restoring and filling the missing track information, and finally outputting the complete track. Specifically, transformer Block blocks in the module adopt a global self-attention mechanism to the input characteristics, and decode to generate a high-dimensional characteristic representation sequence of the complete track; secondly, the MLP component performs dimension reduction processing on the generated complete track high-dimensional characteristic representation sequence to map to a low-dimensional characteristic vector space and then perform characteristic recombination, and finally outputs the complete track sequence. Because the model designs a separate encoding and decoding module filling frame with asymmetric size, quick track filling with high deletion rate can be finally realized.
In step S4, the model for filling the missing track learns the difference between the output complete track and the real complete track by using the loss function, so as to optimize the model parameters, and the calculation formula of the loss function is as follows:
，
wherein ,for the true value of the j-th element of the i-th track point,/for the j-th element>The model output value of the j element of the ith track point is given, the elements refer to longitude, latitude and altitude, H is the total number of track points taken by the track missing filling model according to a missing mode, M is a positive integer, and M is the number of elements of each track point and is constant 3; training a model by adopting a self-supervision training method, in the training process, manually masking the complete track in advance, inputting the missing track after the manual masking into the model for training, finally comparing the complete track output by the model with the complete track without masking, and returning the gradient training model by using an Adam optimizer; in addition, different models are trained by adjusting super parameters and setting different loss rates so as to obtain a track filling model, after training is finished, model verification is continuously carried out on a test set, and finally an optimal track filling model is selected.
The super-parameters of the track loss filling model are as follows:
track embedding and lossless feature coding module with missing mark
(1) Data embedding: carrying out dimension change (patched) on input data Y from (C, H, W) to (L, D), namely numerically CHW=LD;
(2) Adding position codes: adding a position pos emped, using a sin-cos coding method, and avoiding updating position coding parameters in the training process without back propagation;
(3) Adding cls token: the original cls token is added with position codes to obtain clstoken with positions, and the cls token is spliced and added into x according to a post-expansion concat tensor of the batch_size, and the cls token is fully called: class token, chinese translation into class token;
(4) Basic parameters: using Transformer Block as the base coding unit, the patch size is 1, the embedding dimension (encoder_end_dim) is 512,Transformer Block blocks (encoder_depth) is 12, the multi-headed attention awareness (encoder_num_heads) is 12, and the mlp hidden layer node number (mlp _ratio) is 4.
Asymmetric track feature decoding module
(1) Data embedding: inputting x as a decoder enabling the linear network to be used for linear dimension transformation of data;
(2) Mask manufacturing: obtaining a mask position matrix, and generating a corresponding feature matrix which is all 1;
(3) Restoring the Mask dimension: adding mask token concat corresponding dimensions into the input x, recovering the position by using a position matrix, and finally adding an original cls token with the input x again;
(4) Adding position codes: the sin-cos code is fixed to be added with the characteristic as an element, and the position code is not back propagated;
(5) Basic parameters: using Transformer Block blocks as the basic decoding unit, the embedding dimension (decoder_end_dim) is 256, the number of transform block blocks (decoder_depth) is 8, the multi-head attention awareness number (decoder_num_heads) is 16, and the number of mlp network hidden layer nodes (mlp _ratio) is 4.
According to the embodiment, four different artificial mask modes are correspondingly designed according to different track missing conditions, and tracks in different mask modes are alternately input during training to enable robustness of a model to be improved as much as possible, so that the model can have good filling accuracy for various missing conditions in real conditions. Fig. 7 to 10 are four artificial mask patterns designed according to the embodiments, and correspond to the trace missing situations existing in the processes of flat flight, take-off and landing in random and three flight phases respectively, and the points in fig. 7 to 10 are trace points (trace points) representing the positions of the aircraft at a certain moment, and are composed of longitudes (longitudes), latitudes (latitudes) and altitudes (heights), that is, three attributes (longitudes, latitudes and altitudes) are included in one trace Point, and the trace points are visualized into a three-dimensional coordinate system, and all trace points are visualized into the three-dimensional coordinate system, so that the illustration contents of fig. 7 to 10 are formed, and it is noted that the mask points of all mask situations of the present invention are larger than 50%, that is, the missing trace points are larger than the known trace points. The four mask patterns need to satisfy the following pseudocode:
if(c/A<30%)
{mode = random_mode;}
else
{
if(missing_stage == level_flight) {mode = level_model;}
if(missing_stage == take_off) {mode = takeoff_model;}
if(missing_stage == landing) {mode = landing_model;}
}
wherein c represents the number of continuously missing track points, and A represents the total track points.
In step S4, as shown in fig. 11, a schematic diagram of a random missing track is shown, the obtained missing track sequence_dot is input into a trained deep learning model, the model outputs the complete track, and the filling track points at corresponding positions are replaced by true value track points dot_1, dot_2, dot_4 …, dot_n-1 and dot_n existing in the sequence_dot, so that the complete track only filling the missing track is formed by the track, and the specific filling process is shown in fig. 12, wherein the missing track shares n track points, and a number of valid track points are shown in fig. 12. Each track point includes three data, namely longitude, latitude, and altitude, expressed mathematically as (x, y, z). The track is formed by a series of track points arranged in time sequence, and the content in fig. 12 is as follows:
a. counting a number of track points with data from the acquired missing track sequence, deleting invalid data, and embedding tracks;
b. inputting the formed track data vector into a lossless track feature coding module, and extracting features of the track;
c. the encoder performs feature encoding on the input track data vector to generate a high-dimensional feature vector;
d. filling a blank space occupation value (Null) in the high-dimensional feature vector according to the position corresponding to the invalid data in the step a, and adding the blank space occupation value with the position code as an element;
e. inputting the feature vector filling the blank space occupation value into an asymmetric track feature decoding module, and decoding and dimension reducing the feature vector;
f. the asymmetric track characteristic decoding module outputs the decoded full-segment track and reconstructs the full-segment track into a complete track sequence;
g. and finishing track filling.
A processor of the aircraft track loss filling device adopts a Core i7-12700 processor, and a memory adopts a solid state disk of three-star 980 PRO 1T.
The foregoing description of the preferred embodiments of the invention is not intended to be limiting, but rather is intended to cover all modifications, equivalents, and alternatives falling within the spirit and principles of the invention.
Claims (9)
1. The filling method for the aircraft track Gao Queshi rate is characterized by comprising the following steps of:
s1, extracting track data comprising time, longitude, latitude and altitude;
s2, normalizing the extracted track data;
s3, matching the normalized track data with the missing modes, selecting the track data in each missing mode as a track data set according to the same proportion after matching is completed, and dividing the track data set into a training set, a test set and a verification set according to the proportion;
s4, inputting a missing track into the track missing filling model, outputting filled track data by the track missing filling model, and determining the track missing filling model by the following steps: and constructing a track loss filling model comprising a track embedding module with a loss identification, a lossless feature encoding module and an asymmetric track feature decoding module, and training the track loss filling model by using the training set, the testing set and the verification set in the step S3.
2. The method according to claim 1, wherein in step S1, when extracting track data including time, longitude, latitude, and altitude, track message data is extracted from the ATC system, the track message data is decoded, and the decoded data is subjected to data cleaning.
3. The method of claim 1, wherein in step S3, the missing modes include a take-off missing mode, a flat flight missing mode, a landing missing mode, and a random missing mode.
4. A method of filling up an aircraft track Gao Queshi according to claim 3, wherein in step S4, the track embedding module with missing marks includes a block embedding component and a missing position encoding component, the block embedding component receives the missing track sequence X and organizes each known track point therein into a track matrix T according to time sequence, the track matrix T is converted into a column vector V of one dimension by tensor flattening operation and is input to the full link layer, and the missing track sequence X is converted into a high-dimensional missing track sequenceThe missing position coding component adds different coding information to the known and unknown track points respectively, the missing position coding component comprises a missing mark coder and a fixed position coder, and the missing position coder gives a high-dimensional missing track sequence->The track points in the track are added with one-dimensional missing mark codes, and the codes are described as follows by a formula:
，
wherein ,to add a track representation encoded with one-dimensional missing identifications,
，/>for a vector space with one dimension (L, (d+1)), d represents the high-dimensional missing track sequence +.>L is the high-dimensional missing track sequence +.>D and L are positive integers,
fixed position encoder for high-dimensional missing track sequenceAdding fixed position codes to track points in the track points, and describing the fixed position codes by a formula as follows:
，
wherein ,representing a fixed position coding matrix->Element value on row i, row 2j, will +> and />Adding, described by the formula:
，
wherein P is a high-dimensional track vector added with a position coding component,is->And a fixed position coding matrix with consistent dimension.
5. The method of claim 4, wherein in step S4, the lossless feature code module includes N Transformer Block, n+.2, and is an integer, for learning the high-dimensional feature representation of the track point from the high-dimensional track vector P, and the formula is:
，
wherein E represents a high-dimensional feature sequence learned from the high-dimensional track vector P;
transformer Block includes a multi-head group attention component, a multi-layer sensor, and 2-layer normalization.
6. The method of claim 5, wherein in step S4, E is added to the fixed position encoded element and the result is transmitted to the asymmetrical track feature decoding module, expressed as:
，
wherein ,representing a high-dimensional feature sequence with E added to the fixed-position coded feature.
7. The method of filling up an aircraft track Gao Queshi rate according to claim 6, wherein in step S4, the asymmetric track feature decoding module and the lossless track feature encoding module are independent of each other and have different internal neural network parameter sizes and module sizes, the asymmetric track feature decoding module includes 1 Transformer Block and 1 multi-layer sensor, transformer Block pairs ofDecoding to generate a high-dimensional characteristic representation sequence of a complete track by adopting a global self-attention mechanism; the multi-layer perceptron performs dimension reduction processing on the generated complete track high-dimensional characteristic representation sequence to map to a low-dimensional characteristic vector space and then perform characteristic recombination, and finally outputs the complete track sequence.
8. The method of claim 7, wherein in step S4, model training is performed by using a loss function, a self-supervision training method and an Adam optimizer, the loss function is calculated by using different track points in different missing modes, and the calculation formula of the loss function is as follows:
，
wherein ,for the true value of the j-th element of the i-th track point,/for the j-th element>The model output value of the j element of the ith track point is represented by longitude, latitude and height, H is the total number of track points taken by the track missing filling model according to a missing mode, the positive integer is represented by M, the number of elements of each track point is represented by constant 3.
9. A padding apparatus for aircraft track Gao Queshi, comprising at least one processor and at least one memory communicatively coupled to the processor, the memory storing instructions for execution by the at least one processor to enable the at least one processor to perform any of the steps of the method of any of claims 1-8.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202311154958.1A CN116910919B (en) | 2023-09-08 | 2023-09-08 | Filling method and device under Gao Queshi rate of aircraft track |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202311154958.1A CN116910919B (en) | 2023-09-08 | 2023-09-08 | Filling method and device under Gao Queshi rate of aircraft track |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN116910919A true CN116910919A (en) | 2023-10-20 |
| CN116910919B CN116910919B (en) | 2023-12-12 |
Family
ID=88360525
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202311154958.1A Active CN116910919B (en) | 2023-09-08 | 2023-09-08 | Filling method and device under Gao Queshi rate of aircraft track |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN116910919B (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117255363A (en) * | 2023-11-20 | 2023-12-19 | 北京航空航天大学 | Method and system for detecting, repairing and filling defects of signaling data |
Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113408392A (en) * | 2021-06-15 | 2021-09-17 | 西安电子科技大学 | Flight path completion method based on Kalman filtering and neural network |
| CN114595427A (en) * | 2022-03-10 | 2022-06-07 | 华南理工大学 | Filling and repairing method for vehicle track sequence containing missing value based on non-autoregressive model |
| CN115130750A (en) * | 2022-06-28 | 2022-09-30 | 北京信息科技大学 | Flight path prediction method and device, electronic equipment and readable storage medium |
| CN115310677A (en) * | 2022-07-18 | 2022-11-08 | 四川大学 | Flight path prediction method and device based on binary coded representation and multi-classification |
| CN116502777A (en) * | 2023-06-28 | 2023-07-28 | 四川大学 | Transformer-based four-dimensional track long-time prediction method and device |
| CN116543603A (en) * | 2023-07-07 | 2023-08-04 | 四川大学 | Flight path completion prediction method and device considering airspace situation and local optimization |
-
2023
- 2023-09-08 CN CN202311154958.1A patent/CN116910919B/en active Active
Patent Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113408392A (en) * | 2021-06-15 | 2021-09-17 | 西安电子科技大学 | Flight path completion method based on Kalman filtering and neural network |
| CN114595427A (en) * | 2022-03-10 | 2022-06-07 | 华南理工大学 | Filling and repairing method for vehicle track sequence containing missing value based on non-autoregressive model |
| CN115130750A (en) * | 2022-06-28 | 2022-09-30 | 北京信息科技大学 | Flight path prediction method and device, electronic equipment and readable storage medium |
| CN115310677A (en) * | 2022-07-18 | 2022-11-08 | 四川大学 | Flight path prediction method and device based on binary coded representation and multi-classification |
| CN116502777A (en) * | 2023-06-28 | 2023-07-28 | 四川大学 | Transformer-based four-dimensional track long-time prediction method and device |
| CN116543603A (en) * | 2023-07-07 | 2023-08-04 | 四川大学 | Flight path completion prediction method and device considering airspace situation and local optimization |
Non-Patent Citations (2)
| Title |
|---|
| ZHENG ZHANG 等: "Flight trajectory prediction enabled by time-frequency wavelet transform", 《NATURE COMMUNICATIONS》, vol. 14, pages 1 - 15 * |
| 李可欣 等: "基于Transformer＿LSTM编解码器模型的船舶轨迹异常检测方法", 《中国舰船研究》, pages 1 - 11 * |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117255363A (en) * | 2023-11-20 | 2023-12-19 | 北京航空航天大学 | Method and system for detecting, repairing and filling defects of signaling data |
| CN117255363B (en) * | 2023-11-20 | 2024-01-23 | 北京航空航天大学 | Method and system for detecting, repairing and filling defects of signaling data |
Also Published As
| Publication number | Publication date |
|---|---|
| CN116910919B (en) | 2023-12-12 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN113936339B (en) | Fighting identification method and device based on double-channel cross attention mechanism | |
| CN109919031B (en) | Human behavior recognition method based on deep neural network | |
| CN116168352B (en) | Power grid obstacle recognition processing method and system based on image processing | |
| CN116910919B (en) | Filling method and device under Gao Queshi rate of aircraft track | |
| CN110427989B (en) | Automatic synthesis method of Chinese character skeleton and automatic generation method of large-scale Chinese character library | |
| CN112365434B (en) | Unmanned aerial vehicle narrow passage detection method based on double-mask image segmentation | |
| JP7226855B2 (en) | Calculation method, system and storage medium for train passing through tunnel | |
| WO2021013190A1 (en) | Meteorological parameter-based high-speed train positioning method and system in navigation blind zone | |
| CN111832615A (en) | Sample expansion method and system based on foreground and background feature fusion | |
| CN110969107A (en) | Bird population identification analysis method and system based on network model | |
| CN114360030A (en) | Face recognition method based on convolutional neural network | |
| CN111461426A (en) | High-precision travel time length prediction method based on deep learning | |
| CN115374375A (en) | Track clustering method, system and storage medium based on deep learning | |
| CN115752708A (en) | Airport single-point noise prediction method based on deep time convolution network | |
| CN115099328A (en) | Traffic flow prediction method, system, device and storage medium based on countermeasure network | |
| CN114821299A (en) | Remote sensing image change detection method | |
| CN116502777B (en) | Transformer-based four-dimensional track long-time prediction method and device | |
| CN116776014B (en) | Multi-source track data representation method and device | |
| CN117111107A (en) | GNSS fault signal detection and elimination method, system, equipment and medium | |
| CN117058476A (en) | Target detection method based on random uncertainty | |
| CN116469029A (en) | Oilfield on-site violation behavior identification method based on multi-layer perceptron network | |
| CN115544613A (en) | Multi-mode data-driven urban road layout design automation method | |
| CN115147432A (en) | First arrival picking method based on depth residual semantic segmentation network | |
| CN114898215A (en) | Automatic arrangement method of sound barrier | |
| CN113205685B (en) | Short-term traffic flow prediction method based on global-local residual error combination model |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |