Aircraft fraud detection method and device based on multi-view network

Through the aircraft fraud detection method based on multi-view network, using sliding window sampling and multi-view feature analysis network, the problem of ADS-B system being vulnerable to fraud attacks is solved, and the accurate prediction of aircraft types and fraud detection is realized, improving the effectiveness of the detection of fraudulent behaviors. The safety of the air traffic management system.

CN119028180APending Publication Date: 2024-11-26SHENYANG AIRCRAFT DESIGN INST AVIATION IND CORP OF CHINA +1
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202410929182.4
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-11
Publication Date
2024-11-26

AI Technical Summary

Technical Problem

Due to its open sharing characteristics and lack of data encryption and authentication, ADS-B systems are vulnerable to fraudulent attacks, such as false tracking messages from drones, making it difficult for air traffic management systems to distinguish between authorized and unauthorized entities.

Method used

Using aircraft fraud detection method based on multi-view network, the takeoff and cruise data of the aircraft are sampled through sliding windows, latitude and longitude coordinates are extracted, pre-trained multi-view feature analysis network is input, probability sequences are output to characterize aircraft types, and confidence is calculated. to detect whether the aircraft has fraudulent behavior.

Benefits of technology

Accurately detect spoofing attacks in complex flight trajectory data, with high prediction accuracy, can effectively distinguish authorized and unauthorized aircraft, and improve the safety of the air traffic management system.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119028180A_ABST
    Figure CN119028180A_ABST
Patent Text Reader

Abstract

The invention belongs to the technical field of data processing, and particularly relates to an aircraft fraud detection method and device based on a multi-view network, and the method comprises the steps: S1, sampling the take-off data of an aircraft through a sliding window, and randomly sampling a plurality of pieces of cruise data; s2, for each sliding window, extracting longitude and latitude coordinates based on the cruise data, and obtaining a square map covering all the longitude and latitude coordinates in the sliding window; s3, inputting the take-off data, the cruise data of each sliding window and the square map into a pre-trained multi-view feature analysis network, and outputting a probability sequence for representing an aircraft type; s4, calculating the confidence coefficient of the probability sequence; and S5, selecting a plurality of probability sequences with high confidence ranking and corresponding aircraft types, and detecting whether the aircraft has a fraud behavior or not according to the aircraft type with the highest proportion quantity. According to the method, the spoofing attack can be accurately detected in the complex flight path data, and the prediction precision is relatively high.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present application belongs to the field of data processing technology, and in particular relates to an aircraft fraud detection method and device based on a multi-view network. Background Art

[0002] As airspace density increases significantly, traditional surveillance technology will find it increasingly difficult to meet the needs of future air traffic management (ATM) systems. In contrast, ADS-B technology has become an important part of the next generation air transportation system due to its high accuracy, large coverage, support for data sharing and air surveillance.

[0003] ADS-B is the abbreviation of Automatic Dependent Surveillance-Broadcast. It is an avionics device used to provide real-time aircraft position and other related information. The ADS-B system receives position and status data from aircraft through satellite navigation and ground stations, and broadcasts this data to surrounding aircraft and ground stations, thereby achieving real-time monitoring and tracking of aircraft.

[0004] The aircraft is equipped with ADS-B transmitting equipment, which can regularly broadcast the aircraft's identification, location, speed, altitude and other information. This information can be received by other aircraft and ground stations for real-time monitoring of air traffic, provision of air traffic management services and improved flight safety. Another system with similar functions to ADS-B is FLARM, a proprietary electronic system used to provide collision avoidance warnings and information exchange between aircraft. Its full name is Flight Alarm, which means flight alarm system. The FLARM system enables mutual perception and communication between aircraft through wireless communication between equipment installed on the aircraft and ground stations. Each aircraft equipped with FLARM equipment can send and receive the location and alarm information of other aircraft, thereby helping pilots avoid mid-air collisions.

[0005] However, the widespread application of data-driven related technologies in fields such as cloud computing, the Internet of Things, service recommendations, and blockchain has provided attackers with powerful software and hardware support and diverse means of attack, which has gradually weakened the technical advantages of the aviation industry in communication protection. Due to the open and shared nature of the ADS-B protocol, its security faces severe challenges. Specifically, the protocol does not provide data encryption and authentication, and messages are broadcast in a simple and open format, which is extremely vulnerable to eavesdropping, interference, modification, and injection. In addition, authorized aircraft and air traffic control (ATC) stations do not perform identity authentication before sending ADS-B messages, making it impossible for the protocol to distinguish between authorized and unauthorized entities. These factors make the ADS-B system very vulnerable to various fraud attacks. For example, at the 2023 Rugby World Cup held in Toulouse, attackers were able to use drones to target the stadium. During the flight, the drone sent out false tracking messages with the intention of convincing air surveillance personnel that it was a French emergency helicopter. Summary of the invention

[0006] In order to solve the above problems, the present application provides an aircraft fraud detection method and device based on a multi-view network to solve the problem of aircraft fraud through broadcasting ADS-B.

[0007] The first aspect of the present application provides an aircraft fraud detection method based on a multi-view network, which mainly includes:

[0008] Step S1, using a sliding window to sample the takeoff data of an aircraft and randomly sampling a plurality of cruise data, wherein the takeoff data and the cruise data are of the same data type and are derived from the ADS-B database of the aircraft;

[0009] Step S2: for each sliding window, extract the longitude and latitude coordinates based on the cruise data, and intercept a square map covering all the longitude and latitude coordinates in the sliding window from the map system;

[0010] Step S3, inputting the takeoff data, the cruise data of each sliding window and the square map into a pre-trained multi-view feature analysis network, and outputting a probability sequence for characterizing the aircraft type;

[0011] Step S4, calculating the confidence of the probability sequence;

[0012] Step S5: select multiple probability sequences with high confidence rankings and the corresponding aircraft types, and compare the aircraft type with the highest proportion with the aircraft type broadcast by the aircraft through ADS-B to detect whether the aircraft has fraudulent behavior.

[0013] Preferably, in step S1, the sliding window is 128 seconds, and data is collected every 1 second.

[0014] Preferably, step S1 further comprises:

[0015] For takeoff data and cruise data, some types of data are extracted from the ADS-B database according to the preset extraction model as the first type of data;

[0016] For ground stations that obtain aircraft types through ADS-B, multiple airport coordinate data associated with the ground station with trajectories are used as the second type of data;

[0017] The first type of data is concatenated with the second type of data to form the final cruise data that is input into the multi-view feature analysis network.

[0018] Preferably, step S3 further comprises:

[0019] Step S31: for takeoff data and cruise data, respectively use the first network module and the second network module trained by the parallel optimal feature integration strategy to output the first intermediate parameter and the second intermediate parameter; for square maps, use the third network module trained by the lifting sampling feature segmentation strategy to output the third intermediate parameter;

[0020] Step S32: input the three types of intermediate parameters together into a fourth network module trained by a parallel optimal feature integration method to output the probability sequence.

[0021] Preferably, in step S4, the confidence is calculated by the following formula:

[0022]

[0023] Among them, p m is the highest probability value in the probability sequence, p o are the probability values ​​other than the highest probability value in the probability sequence, and N is the number of probability values ​​in the probability sequence.

[0024] The second aspect of the present application is an aircraft fraud detection device based on a multi-view network, characterized in that it includes:

[0025] A flight data acquisition module, used to use a sliding window to sample the takeoff data of the aircraft and randomly sample a plurality of cruise data, wherein the takeoff data and the cruise data are of the same data type and are derived from the ADS-B database of the aircraft;

[0026] A map data acquisition module, for extracting the longitude and latitude coordinates of each sliding window based on the cruise data, and intercepting a square map covering all the longitude and latitude coordinates in the sliding window from the map system;

[0027] A probability sequence generation module is used to input the takeoff data, the cruise data of each sliding window and the square map into a pre-trained multi-view feature analysis network, and output a probability sequence for characterizing the aircraft type;

[0028] A confidence calculation module is used to calculate the confidence of the probability sequence;

[0029] The fraud detection module is used to select multiple probability sequences with high confidence rankings and the corresponding aircraft types, and compare the aircraft type with the aircraft type broadcast by the aircraft through ADS-B to detect whether the aircraft has fraudulent behavior.

[0030] Preferably, the sliding window is 128 seconds, and data is collected every 1 second.

[0031] Preferably, the flight data acquisition module includes:

[0032] An ADS-B data selection unit is used to extract some types of data from the ADS-B database for takeoff data and cruise data according to a preset extraction model as first-category data;

[0033] an airport data acquisition unit, for a ground station that acquires the aircraft type through ADS-B, using a plurality of airport coordinate data associated with the ground station with a trajectory as the second type of data;

[0034] The data splicing unit is used to splice the first type of data with the second type of data to form the final cruise data input into the multi-view feature analysis network.

[0035] Preferably, the probability sequence generation module includes:

[0036] The intermediate layer output unit is used to output the first intermediate parameter and the second intermediate parameter using the first network module and the second network module trained by the parallel optimal feature integration strategy for the takeoff data and the cruise data, respectively, and to output the third intermediate parameter using the third network module trained by the lifting sampling feature segmentation strategy for the square map;

[0037] The probability sequence output unit is used to input the three types of intermediate parameters together into the fourth network module trained by the parallel optimal feature integration method to output the probability sequence.

[0038] Preferably, in the confidence calculation module, the confidence is calculated by the following formula:

[0039]

[0040] Among them, p m is the highest probability value in the probability sequence, p o are the probability values ​​other than the highest probability value in the probability sequence, and N is the number of probability values ​​in the probability sequence.

[0041] This application can predict the aircraft type through the flight trajectory data broadcast by ADS-B, and compare it with the aircraft type broadcast by ADS-B to determine whether the aircraft is fraudulent. This application can accurately detect spoofing attacks in complex flight trajectory data and has high prediction accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 It is a flowchart of a preferred embodiment of the aircraft fraud detection method based on multi-view network of the present application.

[0043] Figure 2 This application Figure 1 Schematic diagram of the parallel optimal feature integration strategy of the illustrated embodiment. DETAILED DESCRIPTION

[0044] In order to make the purpose, technical scheme and advantages of the implementation of this application clearer, the technical scheme in the implementation of this application will be described in more detail in combination with the drawings in the implementation of this application. In the drawings, the same or similar reference numerals throughout represent the same or similar elements or elements with the same or similar functions. The described implementation is a part of the implementation of this application, not all of the implementations. The implementation described below with reference to the drawings is exemplary and is intended to be used to explain this application, and cannot be understood as a limitation on this application. Based on the implementation in this application, all other implementations obtained by ordinary technicians in this field without making creative work are within the scope of protection of this application. The implementation of this application is described in detail below in combination with the drawings.

[0045] The first aspect of the present application provides an aircraft fraud detection method based on a multi-view network, such as Figure 1 As shown, it mainly includes:

[0046] Step S1, using a sliding window to sample the takeoff data of an aircraft and randomly sampling a plurality of cruise data, wherein the takeoff data and the cruise data are of the same data type and are derived from the ADS-B database of the aircraft;

[0047] Step S2: for each sliding window, extract the longitude and latitude coordinates based on the cruise data, and intercept a square map covering all the longitude and latitude coordinates in the sliding window from the map system;

[0048] Step S3, inputting the takeoff data, the cruise data of each sliding window and the square map into a pre-trained multi-view feature analysis network, and outputting a probability sequence for characterizing the aircraft type;

[0049] Step S4, calculating the confidence of the probability sequence;

[0050] Step S5: select multiple probability sequences with high confidence rankings and the corresponding aircraft types, and compare the aircraft type with the highest proportion with the aircraft type broadcast by the aircraft through ADS-B to detect whether the aircraft has fraudulent behavior.

[0051] The present application obtains a number of data through steps S1 and S2, and then in step S3, a probability prediction of the aircraft type is output by a pre-trained multi-view feature analysis network, and the final aircraft type is determined by confidence to be compared with the aircraft type sent by the aircraft through ADS-B broadcast to determine whether the aircraft is fraudulent.

[0052] The data obtained in step S1 and step S2 are used for the multi-view feature analysis network in step S3. Prior to this, the multi-view feature analysis network needs to be trained. The training samples used in the training process are described in detail below. Those skilled in the art can understand that the sample data used in training is essentially the same as the data type and size used in step S1 and step S2.

[0053] The multi-view feature analysis network mainly includes four modules. The first module T1 is used to process takeoff data. The second module T2 is used to process cruise data, which is consistent with the data type in step S1. The third module T3 is used to process map data, which is consistent with the data type in step S2. The fourth module T4 is used to integrate the data processed by the above three modules.

[0054] The first module T1 and the second module T2 use a parallel optimal feature integration strategy to process aircraft cruise data and takeoff data. The parallel optimal feature integration method allows the network to identify key information in the flight trajectory and enhance the extraction of this information. First, the flight trajectory of each cruise phase is stored in a csv file for ADS-B broadcast. Subsequently, the present invention uses a sliding window to randomly sample different aircraft cruise data. In some optional embodiments, in step S1, the sliding window is 128 seconds, and data is collected every 1 second, that is, it corresponds to approximately 2 minutes of flight data, thereby obtaining a short segment of continuous aircraft cruise data. In order to ensure the consistency of data sampling, this embodiment selects a standard time step of 1 second for sampling, and uses timestamps to organize data.

[0055] The type of takeoff data is consistent with that of cruise data, but takeoff data is more representative, and these values ​​can provide additional information for the model. Therefore, this application incorporates takeoff data into the model to improve the accuracy and reliability of detection.

[0056] It should also be noted that in step S1, not all data in the ADS-B database is used, and in step S3, the data transmitted to the network is not just the data in the ADS-B database. For example, in some optional embodiments, step S1 further includes:

[0057] Step S11: extracting some types of data from the ADS-B database for takeoff data and cruise data according to a preset extraction model as first type data;

[0058] Step S12: for a ground station that obtains the aircraft type through ADS-B, multiple airport coordinate data associated with the ground station with a trajectory are used as the second type of data;

[0059] Step S13: concatenate the first type of data with the second type of data to form final cruise data that is input into the multi-view feature analysis network.

[0060] In this embodiment, in step S11, certain features are deleted from the ADS-B database according to preset screening conditions. Features such as "alert", "spi" (special position indicator) and "squawk" (transponder code) are irrelevant to identifying trajectory type anomalies. In addition, the purpose of this application is to predict the aircraft type through a network model, and the aircraft type is usually recorded in the ICAO24 code in the ADS-B database, which is the source of aircraft fraud and is easily tampered with. Therefore, this application only uses the 24-bit address of the International Civil Aviation Organization (ICAO24) as a later comparison item, and will not be used as an input item for the network model.

[0061] Removing features that are considered relatively unimportant for anomaly detection saves computational resources, which enables a more targeted analysis of key features of the task. Table 1 details the various features retained in the dataset.

[0062] Table 1 Dataset characteristics

[0063]

[0064]

[0065] Then in step S12, in addition to the 8 feature data retained above, multiple airport coordinate data associated with the ground station's trajectory are also obtained. For example, if the ground station is Toulouse Airport, 33 airport coordinate data related to the trajectory of the aircraft flying to Toulouse Airport can also be obtained, thereby constructing a 41-dimensional array. In this way, in step S3, a batch or half a batch of 41-dimensional arrays can be input at one time. Here, the batch corresponds to a sliding window. As mentioned above, a sliding window obtains 128 element groups, which can be input into the network model at one time or divided into two inputs into the network model, that is, a 64*41 two-dimensional array is input each time, and each batch will predict a probability sequence.

[0066] refer to Figure 2 , there are multiple data in the obtained 41-dimensional array, some of which are of high importance and some of which are of low importance. The parallel optimal feature integration strategy used by module T1 and the second module T2 refers to assigning different weights to these data. The basic principle is to decompose the input data into different subspaces, apply the optimal feature integration strategy independently in each subspace, and then merge the results of these subspaces to obtain a comprehensive final effect. Specifically, given a retrieval vector S (here refers to selecting a feature value from the 41-dimensional array), and a set of related background vectors {b1, b2, …, b n} (here refers to the remaining 40 eigenvalues ​​in the 41-dimensional array) and a set of value vectors {v1, v2, …, v n}. Then calculate the index vector S and background vector b i Similarity between them, and assign weight W to each eigenvalue in the 41-dimensional array according to the similarity i The formula is as follows:

[0067] Similarity=S*b i ;

[0068]

[0069] The above formula for assigning weights is essentially a softmax normalization function, which is used to convert a set of similarities into a probability distribution so that the sum of all weights is 1, thereby enabling effective weighted averaging. After obtaining the weights, they can be multiplied by the value vector to obtain the final effect value Effect:

[0070]

[0071] The parallel optimal feature integration strategy enables the model to capture different aspects of the input data in different representation subspaces, enhancing the model's ability to process complex data. It also allows the model to focus on different parts of the input sequence at the same time, which is particularly important for the nonlinear and high-dimensional characteristics of flight data. In this way, the model not only focuses on the current state, but also takes into account historical information and its impact on the future.

[0072] The third module for T3 uses an upsampling feature segmentation strategy that processes and analyzes geographic context. The encoder-decoder network of the upsampling feature segmentation strategy is well suited for extracting spatial features from high-resolution geographic images. By gradually extracting and amplifying image features, it ensures that geographic information is preserved during feature extraction. The structure of the autoencoder is the inspiration for the upsampling feature segmentation strategy, which is a convolutional neural network commonly used for image segmentation. The core principle of the upsampling feature segmentation strategy is to capture the content information of the image through a series of convolutional layers and downsampling operations. Precise positioning is then achieved through symmetrical upsampling and convolution operations, resulting in a segmentation map of the same size as the input image.

[0073] The formula is:

[0074] C l =max(0,W l *X l-1 +b l )

[0075] P l =MaxPool(C l ).

[0076] Among them, C l Represents the feature map after the convolution operation, P l is the feature map after the pooling operation, W l and b l are the weights and biases of the convolutional layer, X l-1 is the output of the previous layer, * is the convolution operation, and MaxPool is the maximum pooling operation.

[0077] In step S2, in order to generate a geographic context, a large map image covering the bounding box is generated using an aggregation of OpenStreetMap. After obtaining the aircraft cruise data and takeoff context, the model accurately extracts the area where the latitude and longitude of the aircraft are located within a time window and displays it in a small square of 128x128 pixels. Preferably, the application replaces absolute coordinates with relative coordinates because the map provides absolute positioning information. The use of relative coordinates has the effect of reducing the numerical space required to represent the trajectory. Using relative coordinates, the space required to represent the trajectory is reduced from a rectangle with a diagonal of 155km to 50km. This not only simplifies the learning process, but also encourages the model to prioritize the shape of the trajectory rather than its specific location.

[0078] Based on the above description, the present application uses the first module T1, the second module T2 and the third module T3 to process flight data and geographic data respectively, and combines the up-down sampling feature segmentation strategy with the parallel optimal feature integration strategy to effectively integrate and analyze multi-source data. The entire architecture makes full use of time series data and spatial data. Therefore, the network of the present application achieves high-precision trajectory classification and provides an effective method for abnormal monitoring of aircraft. The output values ​​of the above three modules are further processed by the fourth module T4, and the fourth module T4 also adopts the parallel optimal feature integration strategy. In order to be consistent with the above training process, the present application also uses these four modules in the use of the model in step S3, that is, in some optional embodiments, step S3 further includes:

[0079] Step S31: for takeoff data and cruise data, respectively use the first network module and the second network module trained by the parallel optimal feature integration strategy to output the first intermediate parameter and the second intermediate parameter; for square maps, use the third network module trained by the lifting sampling feature segmentation strategy to output the third intermediate parameter;

[0080] Step S32: input the three types of intermediate parameters together into a fourth network module trained by a parallel optimal feature integration method to output the probability sequence.

[0081] Then, the present invention performs confidence calculation on the output multiple probability sequences in step S4. In some optional implementations, in step S4, the confidence is calculated by the following formula:

[0082]

[0083] Among them, p m is the highest probability value in the probability sequence, p o are the probability values ​​other than the highest probability value in the probability sequence, and N is the number of probability values ​​in the probability sequence.

[0084] For example, suppose the probability sequence contains three types of aircraft, namely business aircraft, tourist aircraft and helicopter. One probability sequence is (0.7, 0.1, 0.2) and the other probability sequence is (0.9, 0.05, 0.05). For the first probability sequence, the category with the highest recognition probability is commercial aircraft, with a probability value of 0.7. The average predicted probability of the other two categories is (0.1+0.2) / 2=0.15, so the confidence score of the first sequence is 0.7-0.15=0.55. For the second probability sequence, the category with the highest recognition probability is commercial aircraft, with a probability value of 0.9, and the average predicted probability of the other two categories is (0.05+0.05) / 2=0.05, so the confidence score of the second trajectory is 0.9-0.05=0.85. According to the definition and calculation method of the confidence, the confidence and reliability of the second sequence are higher than those of the first sequence.

[0085] Finally, in step S5, multiple probability sequences with high confidence rankings and corresponding aircraft types are selected. For example, 20 high-confidence probability sequences are selected, and the aircraft type with the highest proportion is compared with the aircraft type broadcast by the aircraft through ADS-B to detect whether the aircraft is fraudulent. For example, if there are 15 probability sequences indicating that the aircraft is a commercial aircraft, the aircraft type finally output by the model is a commercial aircraft. When the aircraft type broadcast by the aircraft through ADS-B is other non-commercial aircraft, it is clear that the aircraft is fraudulent.

[0086] The second aspect of the present application provides an aircraft fraud detection device based on a multi-view network corresponding to the above method, which mainly includes:

[0087] A flight data acquisition module, used to use a sliding window to sample the takeoff data of the aircraft and randomly sample a plurality of cruise data, wherein the takeoff data and the cruise data are of the same data type and are derived from the ADS-B database of the aircraft;

[0088] A map data acquisition module, for extracting the longitude and latitude coordinates of each sliding window based on the cruise data, and intercepting a square map covering all the longitude and latitude coordinates in the sliding window from the map system;

[0089] A probability sequence generation module is used to input the takeoff data, the cruise data of each sliding window and the square map into a pre-trained multi-view feature analysis network, and output a probability sequence for characterizing the aircraft type;

[0090] A confidence calculation module is used to calculate the confidence of the probability sequence;

[0091] The fraud detection module is used to select multiple probability sequences with high confidence rankings and the corresponding aircraft types, and compare the aircraft type with the aircraft type broadcast by the aircraft through ADS-B to detect whether the aircraft has fraudulent behavior.

[0092] In some optional implementations, the sliding window is 128 seconds, and data is collected every 1 second.

[0093] In some optional implementations, the flight data acquisition module includes:

[0094] An ADS-B data selection unit is used to extract some types of data from the ADS-B database for takeoff data and cruise data according to a preset extraction model as first-category data;

[0095] an airport data acquisition unit, for a ground station that acquires the aircraft type through ADS-B, using a plurality of airport coordinate data associated with the ground station with a trajectory as the second type of data;

[0096] The data splicing unit is used to splice the first type of data with the second type of data to form the final cruise data input into the multi-view feature analysis network.

[0097] In some optional implementations, the probability sequence generation module includes:

[0098] The intermediate layer output unit is used to output the first intermediate parameter and the second intermediate parameter using the first network module and the second network module trained by the parallel optimal feature integration strategy for the takeoff data and the cruise data, respectively, and to output the third intermediate parameter using the third network module trained by the lifting sampling feature segmentation strategy for the square map;

[0099] The probability sequence output unit is used to input the three types of intermediate parameters together into the fourth network module trained by the parallel optimal feature integration method to output the probability sequence.

[0100] In some optional implementations, in the confidence calculation module, the confidence is calculated by the following formula:

[0101]

[0102] Among them, p m is the highest probability value in the probability sequence, p o are the probability values ​​other than the highest probability value in the probability sequence, and N is the number of probability values ​​in the probability sequence.

[0103] The above is only a specific implementation of the present application, but the protection scope of the present application is not limited thereto. Any changes or substitutions that can be easily thought of by a person skilled in the art within the technical scope disclosed in the present application should be included in the protection scope of the present application. Therefore, the protection scope of the present application shall be based on the protection scope of the claims.

Claims

1. A method for detecting aircraft fraud based on a multi-view network, characterized in that: include: Step S1, using a sliding window to sample the takeoff data of an aircraft and randomly sampling a plurality of cruise data, wherein the takeoff data and the cruise data are of the same data type and are derived from the ADS-B database of the aircraft; Step S2: for each sliding window, extract the longitude and latitude coordinates based on the cruise data, and intercept a square map covering all the longitude and latitude coordinates in the sliding window from the map system; Step S3, inputting the takeoff data, the cruise data of each sliding window and the square map into a pre-trained multi-view feature analysis network, and outputting a probability sequence for characterizing the aircraft type; Step S4, calculating the confidence of the probability sequence; Step S5: select multiple probability sequences with high confidence rankings and the corresponding aircraft types, and compare the aircraft type with the highest proportion with the aircraft type broadcast by the aircraft through ADS-B to detect whether the aircraft has fraudulent behavior.

2. The aircraft fraud detection method based on multi-view network as claimed in claim 1, characterized in that: In step S1, the sliding window is 128 seconds, and data is collected every 1 second.

3. The aircraft fraud detection method based on multi-view network as claimed in claim 1, characterized in that: Step S1 further comprises: For takeoff data and cruise data, some types of data are extracted from the ADS-B database according to the preset extraction model as the first type of data; For ground stations that obtain aircraft types through ADS-B, multiple airport coordinate data associated with the ground station with trajectories are used as the second type of data; The first type of data is concatenated with the second type of data to form the final cruise data that is input into the multi-view feature analysis network.

4. The aircraft fraud detection method based on multi-view network as claimed in claim 1, characterized in that: Step S3 further comprises: Step S31: for takeoff data and cruise data, respectively use the first network module and the second network module trained by the parallel optimal feature integration strategy to output the first intermediate parameter and the second intermediate parameter; for square maps, use the third network module trained by the lifting sampling feature segmentation strategy to output the third intermediate parameter; Step S32: input the three types of intermediate parameters together into a fourth network module trained by a parallel optimal feature integration method to output the probability sequence.

5. The aircraft fraud detection method based on multi-view network as claimed in claim 1, characterized in that: In step S4, the confidence is calculated by the following formula: Among them, p m is the highest probability value in the probability sequence, p o are the probability values ​​other than the highest probability value in the probability sequence, and N is the number of probability values ​​in the probability sequence.

6. An aircraft fraud detection device based on a multi-view network, characterized in that: include: A flight data acquisition module, used to use a sliding window to sample the takeoff data of the aircraft and randomly sample a plurality of cruise data, wherein the takeoff data and the cruise data are of the same data type and are derived from the ADS-B database of the aircraft; A map data acquisition module, for extracting the longitude and latitude coordinates of each sliding window based on the cruise data, and intercepting a square map covering all the longitude and latitude coordinates in the sliding window from the map system; A probability sequence generation module is used to input the takeoff data, the cruise data of each sliding window and the square map into a pre-trained multi-view feature analysis network, and output a probability sequence for characterizing the aircraft type; A confidence calculation module is used to calculate the confidence of the probability sequence; The fraud detection module is used to select multiple probability sequences with high confidence rankings and the corresponding aircraft types, and compare the aircraft type with the aircraft type broadcast by the aircraft through ADS-B to detect whether the aircraft has fraudulent behavior.

7. The aircraft fraud detection device based on a multi-view network as claimed in claim 6, characterized in that: The sliding window is 128 seconds, and data is collected every 1 second.

8. The aircraft fraud detection device based on a multi-view network as claimed in claim 6, characterized in that: The flight data acquisition module includes: An ADS-B data selection unit is used to extract some types of data from the ADS-B database for takeoff data and cruise data according to a preset extraction model as first-category data; an airport data acquisition unit, for a ground station that acquires the aircraft type through ADS-B, using a plurality of airport coordinate data associated with the ground station with a trajectory as the second type of data; The data splicing unit is used to splice the first type of data with the second type of data to form the final cruise data input into the multi-view feature analysis network.

9. The aircraft fraud detection device based on a multi-view network as claimed in claim 6, characterized in that: The probability sequence generation module comprises: The intermediate layer output unit is used to output the first intermediate parameter and the second intermediate parameter using the first network module and the second network module trained by the parallel optimal feature integration strategy for the takeoff data and the cruise data, respectively, and to output the third intermediate parameter using the third network module trained by the lifting sampling feature segmentation strategy for the square map; The probability sequence output unit is used to input the three types of intermediate parameters together into the fourth network module trained by the parallel optimal feature integration method to output the probability sequence.

10. The aircraft fraud detection device based on a multi-view network as claimed in claim 6, characterized in that: In the confidence calculation module, the confidence is calculated by the following formula: Among them, p m is the highest probability value in the probability sequence, p o are the probability values ​​other than the highest probability value in the probability sequence, and N is the number of probability values ​​in the probability sequence.