Dual-channel pre-training based multi-target flight trajectory identifier

Abstract

This invention discloses a multi-target flight trajectory recognition device based on dual-channel pre-training, comprising a flight trajectory database, an aircraft radiation source receiver, and a host computer connected to each other. The database stores flight trajectory signals of different types of aircraft. The host computer includes three modules: self-supervised pre-training, supervised training, and flight trajectory recognition. The data in the database is used to perform self-supervised pre-training and supervised training on the flight trajectory recognition model, and the trained model is stored in the flight trajectory recognition module. The receiver receives external flight trajectory signals and transmits them to the flight trajectory recognition module of the host computer for recognition. This invention innovatively introduces a self-supervised pre-training stage to pre-train the flight trajectory recognition model to fully explore its inherent features and enhance its anti-interference capability. Simultaneously, it analyzes the flight trajectory signal from both the time and frequency domains to maximize the extraction of multi-domain and multi-dimensional information from the signal, significantly improving the accuracy of flight trajectory recognition.

Description

Technical Field

[0001] This invention relates to the fields of flight trajectory recognition and data-driven approaches, and in particular, to a multi-target intelligent flight trajectory recognition technology based on self-supervised learning and supervised learning, which is a novel flight trajectory recognition device. Background Technology

[0002] Flight trajectory identification (FTI) is a technology used to determine the type, location, and other characteristics of aircraft. Electromagnetic signals emitted by an aircraft can be captured by a receiver. FTI uses advanced signal processing algorithms to analyze the frequency, time domain, and phase characteristics of the aircraft signals to identify and distinguish the flight trajectories of different aircraft. For example, this technology can identify and locate aircraft interference sources and signal sources, allowing for necessary measures to reduce interference and improve communication quality, thereby enhancing the reliability and security of communication networks. It can also identify and locate radiation sources such as drones, missiles, and satellites, enabling necessary measures to protect equipment and personnel. Furthermore, in electronic warfare, it can be used to identify drones, communication, and navigation equipment. Despite the attention and research FTI has received due to its importance, previous FTI devices still have many significant drawbacks. Previous identification devices were highly sensitive to interference; when a radiation source was interfered with, its emitted signal could be distorted or jammed, leading to a significant decrease in identification accuracy. Simultaneously, this technology requires complex signal processing and analysis, resulting in high computational complexity, long processing times, and low identification accuracy.

[0003] With the increasing maturity of artificial intelligence technology, data-driven technologies have also seen significant development. Data-driven technologies can be divided into supervised learning and unsupervised learning. In supervised learning, the model is trained using training samples with input features and corresponding labels, drawn from known datasets. By learning the mapping from input to output, the model can predict the output of new, unseen data. Unsupervised learning learns the inherent structure and features of data from unlabeled data without requiring any manually labeled output information. In unsupervised learning, the training dataset only contains input data, without output labels. The model needs to automatically discover patterns and structures from the data and perform corresponding learning and processing. Self-supervised learning is a form of unsupervised learning, but unlike traditional unsupervised learning, it uses self-generated labels. In self-supervised learning, data is typically divided into two parts: one as input and the other as the target. The model's learning objective is to predict or understand the target part, thereby learning a useful representation of the data.

[0004] Several applications of data-driven supervised learning methods have been developed for flight trajectory recognition. In these applications, supervised learning typically employs classification algorithms to train models that categorize different types of flight trajectory signals. Specifically, supervised learning requires a labeled training dataset containing various types of flight trajectory signal data and their corresponding labels. A classification algorithm is then designed to train the training dataset, establishing a flight trajectory classification model. When identifying unknown aircraft signal sources, these signals can be input into the model for classification, thereby determining their type and origin. Supervised learning can leverage existing training datasets to improve the accuracy and reliability of recognition and classification, and can also adapt to new flight trajectory signal types and sources. However, supervised learning is also limited by the quality and quantity of the training dataset, requiring a sufficient and representative dataset to achieve accurate recognition and classification. Furthermore, previous algorithms have not been highly optimized and lack sufficient recognition capabilities when complex signal processing and analysis are required, resulting in low flight trajectory recognition accuracy.

[0005] In flight trajectory recognition, self-supervised learning can extract and learn features based on the inherent characteristics and structural information of the signal itself, thus adapting to new signal types and sources and exhibiting better generalization ability. Simultaneously, self-supervised learning can utilize the structural and attribute information of the data itself, thereby enhancing the flight trajectory recognition model's anti-interference capabilities. However, currently, there are few methods and devices that incorporate self-supervised learning to assist in flight trajectory recognition. Previous methods have not fully utilized flight trajectory data for self-supervised pre-training, resulting in insufficient mining of the essential characteristics of the aircraft's radiation source, thus limiting flight trajectory recognition capabilities and causing low instrument recognition accuracy. In conclusion, there is an urgent need to invent better recognition instruments to achieve real-time, high-precision flight trajectory recognition. Summary of the Invention

[0006] In order to overcome the shortcomings of current flight trajectory recognition devices, such as low anti-interference ability, low recognition accuracy, poor feature extraction ability, and low intelligence, the purpose of this invention is to provide a multi-target flight trajectory recognition device with strong feature extraction ability, high accuracy, real-time performance, and high intelligence.

[0007] The technical solution adopted by this invention to solve its technical problem is: a multi-target flight trajectory recognition device based on dual-channel pre-training, including a flight trajectory database, an aircraft radiation source receiver, and a host computer. The flight trajectory database and the aircraft radiation source receiver are connected to the host computer, which includes three modules connected in sequence: a self-supervised pre-training module, a supervised training module, and a flight trajectory recognition module. The recognition device performs flight trajectory recognition according to the following process:

[0008] 1) Label the existing flight trajectory signals with type labels to construct a flight trajectory dataset. The flight trajectory signal dataset is in the following format:

[0009] [(s1,t1),(s2,t2),…,(s n ,t n )]

[0010] One of the parentheses represents a data point of electromagnetic radiation from an aircraft; s represents the time-domain signal of the flight trajectory, which consists of multiple signal amplitudes sampled at a certain time step T, i.e., s = {s 1 ,s 2 ,…,s L}, an s i The amplitude of a sampling point is represented by L, which is the length of the time domain signal of the flight trajectory, i.e., the total number of sampling points; t represents the flight trajectory type corresponding to s; n represents the total number of data; and the total number of flight trajectory categories is K.

[0011] 2) Input the flight trajectory dataset into the flight trajectory database. The flight trajectory data in the database can be used by the host computer as training samples for the flight trajectory recognition model.

[0012] The host computer of the identification device includes:

[0013] The self-supervised pre-training module is used to perform self-supervised pre-training on the flight trajectory recognition model to improve the recognition device's resistance to electromagnetic interference. This is accomplished through the following process:

[0014] 3) Use Fast Fourier Transform to transform the original time-domain flight trajectory signal s in the flight trajectory database into a frequency-domain flight trajectory signal.

[0015]

[0016] FFT stands for Fast Fourier Transform.

[0017] 4) For the time-domain flight trajectory signal s and the frequency-domain flight trajectory signal Both are masked according to a ratio p, that is, the time-domain flight trajectory signal s and the frequency-domain flight trajectory signal are masked. All information is randomly missing in proportion p:

[0018] s m =Mask(s)

[0019]

[0020] Where Mask represents the masking operation, s m The signal after masking the time-domain flight trajectory signal s. Frequency domain flight trajectory signal The signal behind the mask.

[0021] 5) Mask the time-domain flight trajectory signal s m and mask frequency domain flight trajectory signal Simultaneously, the signal is fed into the flight trajectory recognition model for signal reconstruction. The flight trajectory recognition model is a dual-channel neural network N, containing one time-domain channel and one frequency-domain channel. The time-domain channel receives the masked time-domain flight trajectory signal s. m Its self-supervised pre-training objective is to improve s m The original time-domain flight trajectory signal s is reconstructed to fully extract the inherent time-domain characteristics of the flight trajectory signal; the frequency-domain channel receives the masked frequency-domain flight trajectory signal. Its self-supervised pre-training objective is to Reconstructed into the original frequency domain flight trajectory signal To fully exploit the inherent frequency domain characteristics of the flight trajectory signal, the self-supervised pre-training process of this dual-channel neural network N is performed as follows:

[0022] 5.1) N pairs of all masked time-domain flight trajectory signals s m Perform reconstruction and calculate the reconstruction error r;

[0023] Alternatively, as a preferred option, the reconstruction error of the masked time-domain flight trajectory signal can be calculated using the mean square error formula:

[0024]

[0025]

[0026] in Let r be the time-domain flight trajectory signal reconstructed by neural network N, and r be the time-domain reconstruction error.

[0027] 5.2) N pairs of all masked frequency domain flight trajectory signals Perform reconstruction and calculate reconstruction error.

[0028] Alternatively, as a preferred option, the reconstruction error of the masked frequency domain flight trajectory signal can be calculated using the mean square error formula:

[0029]

[0030]

[0031] in The frequency domain flight trajectory signal reconstructed by neural network N. For frequency domain reconstruction error;

[0032] 5.3) Calculate the total reconstruction error The reconstruction error is used to backpropagate the two-channel neural network, and the parameters θ of the two-channel neural network are updated using stochastic gradient descent.

[0033] 5.4) Repeat steps 5.1)-5.3) until the total reconstruction error no longer decreases, and the pre-training of the dual-channel neural network N is completed;

[0034] 5.5) The self-supervised pre-training module saves the dual-channel neural network N and its network parameters θ after pre-training in step 5.4) and transfers them to the supervised training module.

[0035] The supervised training module is used to supervise the training of the flight trajectory recognition model to improve the accuracy of the recognition device. This is accomplished through the following process:

[0036] 6) Use Fast Fourier Transform (FFT) to transform the original time-domain flight trajectory signal s in the flight trajectory database into a frequency-domain flight trajectory signal. Same as step 3), the transformed flight trajectory data is t represents the flight trajectory type label;

[0037] 7) Supervised training is performed on the dual-channel neural network N input from the self-supervised pre-training module. The training process is as follows:

[0038] 7.1) To enable the recognition model to learn the time-domain features of the flight trajectory signal in depth, the time-domain flight trajectory signal s is input into the time-domain channel of the dual-channel neural network N. The time-domain channel extracts the time-domain features from s according to the following calculation process:

[0039] h i =σ(W h s i +v h ⊙c i-1 )

[0040] c i =h i ⊙c i-1 +(1-h i )⊙(W c z i )

[0041] r i =σ(W r z i +v r ⊙c i-1 )

[0042] F t =r i ⊙c i+(1-r i )⊙z i

[0043] Among them W h W c W r Represents the learnable parameter matrix; h i c i r i This represents the intermediate vector generated during the computation process, where i represents the i-th time step, and s... i Let F represent the amplitude of the flight trajectory signal s at the i-th time step, ⊙ represent the dot product, and F t σ represents the time-domain feature vector extracted by the neural network from s; σ represents the Sigmoid function, which is calculated as follows:

[0044]

[0045] 7.2) To enable the recognition model to learn the frequency domain features of the flight trajectory signal in depth, the frequency domain flight trajectory signal is... The frequency domain channels of the input dual-channel neural network N are calculated using the following formula. Extracting frequency domain features from the data:

[0046]

[0047] Among them W q W k W v A and B represent learnable parameter matrices, and the superscript T represents matrix transpose. Introducing matrices A and B can improve the computational efficiency of the model; F f Representing neural networks from The frequency domain feature vector extracted from it; Softmax represents the Softmax function, whose calculation formula is as follows:

[0048]

[0049] 7.3) The time-domain feature vector F obtained in step 7.1) t And the frequency domain eigenvector F obtained in step 7.2) f The features are fed into two adapters for feature transformation and dimensionality reduction respectively:

[0050] O t =Adaptor1(F t )

[0051] O f =Adaptor2(F f )

[0052] Adaptor1 and Adaptor2 represent adapter 1 and adapter 2, respectively, both of which are composed of a single-layer linear neural network; O t Representing F t The time-domain flight trajectory output vector after Adaptor1, O f Representing F f Frequency domain flight trajectory output vector after passing through Adaptor2;

[0053] 7.4) The time-domain flight trajectory output vector O obtained in step 7.3) t and frequency domain flight trajectory output vector O f The data is then concatenated and fed into the linear classification layer of neural network N to classify the flight trajectory signal.

[0054]

[0055] Where: represents vector concatenation operation; Linear represents a linear classification layer; This represents the type of flight trajectory predicted by the neural network.

[0056] 7.5) Calculate the prediction results for all data in the flight trajectory database according to steps 7.1)-7.4) and calculate the total classification error E;

[0057] Alternatively, as a preferred approach, the total classification error E can be calculated using the cross-entropy formula:

[0058]

[0059] 7.6) Backpropagate the training error E to the two-channel neural network and update the parameters θ of the two-channel neural network using stochastic gradient descent.

[0060] 7.7) Repeat steps 7.1)-7.6) until the total classification error no longer decreases, and the training of the dual-channel neural network N is complete;

[0061] 7.8) Input the trained flight trajectory recognition model, i.e., the dual-channel neural network N, into the flight trajectory recognition module.

[0062] The flight trajectory recognition module is used to identify new flight trajectory signals using a flight trajectory recognition model fed from the supervised training module. It can intelligently recognize various types of flight trajectory signals in real time. This is accomplished through the following process:

[0063] 8) The flight trajectory receiver monitors incoming flight trajectory signals in real time and receives new flight trajectory signals s. new Then input it into the flight trajectory recognition module;

[0064] 9) The flight trajectory recognition module will... new The signal is fed into a trained dual-channel neural network N for identification, resulting in a new flight trajectory signal type t. new ;

[0065] 10) The flight trajectory recognition module will display the recognition results and (s) new ,t new The new data sample is input into the flight trajectory database and stored. The stored new flight trajectory data will be used for the reinforcement training of the flight trajectory recognition model in the future.

[0066] The technical concept of this invention is as follows: intelligently identify flight trajectory signals, construct and store flight trajectory datasets, use the datasets to train a dual-channel neural network, establish a flight trajectory recognition model, and obtain a flight trajectory recognition model that meets the task requirements after self-supervised pre-training and supervised training, which is used to identify flight trajectory signals received by the receiver.

[0067] The beneficial effects of this invention are mainly reflected in the following aspects: 1. It innovatively introduces a self-supervised pre-training stage to pre-train the flight trajectory recognition model, enabling the flight trajectory recognition model to fully explore the inherent features of the flight trajectory signal and improve the recognition device's anti-electromagnetic signal interference capability; 2. It comprehensively analyzes the flight trajectory signal from both the time domain and frequency domain, resulting in a high degree of algorithm optimization and the ability to extract multi-domain and multi-dimensional information of the aircraft signal, thus making the recognition device have strong recognition capabilities and high recognition accuracy; 3. The recognition device can identify multiple types of flight trajectory signals in real time, demonstrating strong intelligence. Attached Figure Description

[0068] Figure 1 This is a schematic diagram of the structure of the identification device proposed in this invention;

[0069] Figure 2 This is a schematic diagram of the flight trajectory recognition model proposed in this invention. Detailed Implementation

[0070] The present invention will now be described in detail with reference to the accompanying drawings.

[0071] like Figure 1 This multi-target flight trajectory recognition device based on dual-channel pre-training includes a flight trajectory database 1, a host computer 2, and an aircraft radiation source receiver 6. The flight trajectory database and the aircraft radiation source receiver are connected to the host computer, which includes three modules: a self-supervised pre-training module 3, a supervised training module 4, and a flight trajectory recognition module 5. The recognition device uses the following process to identify underwater targets:

[0072] 1) Collect flight trajectory data according to the required flight trajectory signal type, and label the collected flight trajectory signals with type labels to construct a flight trajectory dataset in the following form:

[0073] [(s 11 ,t1),(s 12 ,t1),…,(s 21 ,t2),(s 22 ,t2),…,(s 81 ,t8),(s 82 ,t8)]

[0074] One of the parentheses represents a single flight trajectory data point; s represents the time-domain signal of the flight trajectory, which consists of multiple signal amplitudes sampled at a time step of T = 2.5 ms, i.e., s = {s 1 ,s 2 ,…,s L}, an s i L represents the amplitude of a sampling point, L = 4000 is the length of the flight trajectory time domain signal, i.e., the total number of sampling points; t represents the flight trajectory type corresponding to s, and in this example, the total number of flight trajectory types is 8;

[0075] 2) Input the flight trajectory dataset into the flight trajectory database. The flight trajectory data in the database can be used by the host computer as training samples for the flight trajectory recognition model.

[0076] The self-supervised pre-training module 3 in the host computer is used to perform self-supervised pre-training on the flight trajectory recognition model. The structure of the flight trajectory recognition model is as follows: Figure 2 As shown, the pre-training process is as follows:

[0077] 3) Use Fast Fourier Transform to transform the original time-domain flight trajectory signal s in the flight trajectory database into a frequency-domain flight trajectory signal.

[0078]

[0079] FFT stands for Fast Fourier Transform.

[0080] 4) For the time-domain flight trajectory signal s and the frequency-domain flight trajectory signal Both are masked according to a ratio p = 0.15, that is, the time-domain flight trajectory signal s and the frequency-domain flight trajectory signal are masked. All of them randomly have 15% of their information missing:

[0081] s m =Mask(s)

[0082]

[0083] Where Mask represents the masking operation, s m The signal after masking the time-domain flight trajectory signal s. Frequency domain flight trajectory signal The signal behind the mask.

[0084] 5) Mask the time-domain flight trajectory signal s m and mask frequency domain flight trajectory signal Simultaneously, the signal is fed into the flight trajectory recognition model for signal reconstruction. The flight trajectory recognition model is as follows: Figure 2 The dual-channel neural network N shown contains one time-domain channel and one frequency-domain channel. The time-domain channel receives the masked time-domain flight trajectory signal s. m Its self-supervised pre-training objective is to improve s m Reconstructed into the original time-domain flight trajectory signal s; frequency-domain channel receives the masked frequency-domain flight trajectory signal. Its self-supervised pre-training objective is to Reconstructed into the original frequency domain flight trajectory signal Previous approaches primarily considered only frequency domain information, neglecting to fully utilize time domain features. This invention, however, extracts features from both the frequency and time domains. The self-supervised pre-training process of this dual-channel neural network N is performed as follows:

[0085] 5.1) N pairs of all masked time-domain flight trajectory signals s m To reconstruct the signal, optionally, the reconstruction error r of the mask time-domain flight trajectory signal can be calculated using the mean square error formula:

[0086]

[0087]

[0088] in Let r be the time-domain flight trajectory signal reconstructed by neural network N, r be the time-domain reconstruction error, and n be the total number of flight trajectory signal samples in the flight trajectory dataset.

[0089] 5.2) N pairs of all masked frequency domain flight trajectory signals To reconstruct the signal, optionally, the reconstruction error of the masked frequency domain flight trajectory signal can be calculated using the mean square error formula.

[0090]

[0091]

[0092] in The frequency domain flight trajectory signal reconstructed by neural network N. The frequency domain reconstruction error is denoted as n, where n is the total number of flight trajectory signal samples in the flight trajectory dataset.

[0093] 5.3) Calculate the total reconstruction error The reconstruction error is used to backpropagate the two-channel neural network, and the parameters θ of the two-channel neural network are updated using stochastic gradient descent.

[0094] 5.4) Repeat steps 5.1)-5.3) until the total reconstruction error no longer decreases. Optionally, an error threshold τ can be set. When the total reconstruction error is below the threshold τ after 3 consecutive iterations, the pre-training of the dual-channel neural network N can be considered complete.

[0095] Optionally, the time-domain and frequency-domain channels of the flight trajectory recognition model can be self-supervised pre-trained separately. That is, the parameters of the frequency-domain channel of the recognition model are first fixed, and only the time-domain flight trajectory signal s is masked. m The signal is fed into the time-domain channel of the recognition model for signal reconstruction. The reconstruction error *r* of the masked time-domain flight trajectory signal is calculated. Based on *r*, the time-domain channel parameters of the recognition model are updated using stochastic gradient descent until the time-domain reconstruction error *r* no longer decreases. Then, the parameters of the pre-trained recognition model's time-domain channels are fixed, and only the masked frequency-domain flight trajectory signal is processed. The signal is fed into the frequency domain channel of the recognition model for signal reconstruction, and the reconstruction error of the masked frequency domain flight trajectory signal is calculated. according to The frequency domain channel parameters of the recognition model are updated using stochastic gradient descent until the frequency domain reconstruction error is reached. No further reduction. This optional approach of performing self-supervised pre-training on the time domain and frequency domain channels of the flight trajectory recognition model separately is because the self-supervised task of reconstructing the flight trajectory signal is quite difficult. Pre-training the channels in both domains at the same time may make it difficult for the recognition model to converge during the pre-training process. However, performing pre-training separately can accelerate the convergence speed of the model and allow the recognition model to obtain the optimal parameters as soon as possible.

[0096] 5.5) The self-supervised pre-training module saves the dual-channel neural network N and its network parameters θ after pre-training in step 5.4) and transfers them to the supervised training module.

[0097] Generally, acquiring flight trajectory data is difficult, constructing a well-calibrated dataset is costly and challenging, and the number of samples in such datasets is typically limited. Current flight trajectory recognition models are often trained using data-driven methods, assuming sufficient data samples, which is often difficult to meet in real-world flight trajectory recognition tasks. A small number of flight trajectory data samples makes it challenging to improve the recognition model's accuracy. This invention employs a self-supervised pre-training module, which fully utilizes the data samples themselves to mine their inherent features. This ensures the recognition model has excellent initialization parameters before supervised training, greatly aiding in model optimization and significantly improving the performance of the recognition device.

[0098] The supervised training module 4 in the host computer is used to train the flight trajectory recognition model, and is completed through the following process:

[0099] 6) Use Fast Fourier Transform (FFT) to transform the original time-domain flight trajectory signal s in the flight trajectory database into a frequency-domain flight trajectory signal. Same as step 3), the transformed flight trajectory data is t represents the flight trajectory type label;

[0100] 7) Supervised training is performed on the dual-channel neural network N input from the self-supervised pre-training module. The training process is as follows:

[0101] 7.1) To enable the recognition model to learn the time-domain features of the flight trajectory signal in depth, the time-domain flight trajectory signal s is input into the time-domain channel of the dual-channel neural network N. The time-domain channel extracts the time-domain features from s according to the following calculation process:

[0102] h i =σ(W h s i +v h ⊙c i-1 )

[0103] c i =h i ⊙c i-1 +(1-h i )⊙(W c z i )

[0104] r i =σ(W r z i +v r ⊙c i-1 )

[0105] F t =r i ⊙c i +(1-r i)⊙z i

[0106] Among them W h W c W r Represents the learnable parameter matrix; h i c i r i This represents the intermediate vector generated during the computation process, where i represents the i-th time step, and s... i Let F represent the amplitude of the flight trajectory signal s at the i-th time step, ⊙ represent the dot product, and F t σ represents the time-domain feature vector extracted by the neural network from s; σ represents the Sigmoid function, which is calculated as follows:

[0107]

[0108] 7.2) To enable the recognition model to learn the frequency domain features of the flight trajectory signal in depth, the frequency domain flight trajectory signal is... The frequency domain channels of the input dual-channel neural network N are calculated using the following formula. Extracting frequency domain features from the data:

[0109]

[0110] Among them W q W k W v A and B represent learnable parameter matrices, and the superscript T represents matrix transpose. Introducing matrices A and B can improve the computational efficiency of the model and reduce the number of parameters required for the recognition model; F f Representing neural networks from The frequency domain feature vector extracted from it; Softmax represents the Softmax function, whose calculation formula is as follows:

[0111]

[0112] Traditional recognition devices manually extract features by performing empirical mode decomposition, Hilbert-Huang transform, and bispectral feature extraction on signals. These manual conversion methods introduce additional computational complexity and limit the model's learning process to predefined features. In contrast, this invention directly utilizes the original time-domain and frequency-domain signals, avoiding the manual feature selection process and enabling the model to freely mine signal features.

[0113] 7.3) The time-domain feature vector F obtained in step 7.1) t And the frequency domain eigenvector F obtained in step 7.2) f The components are fed into two adapters respectively:

[0114] Ot =Adaptor1(F t )

[0115] O f =Adaptor2(F f )

[0116] Here, Adaptor1 and Adaptor2 represent Adapter 1 and Adapter 2, respectively. Both consist of a single-layer linear neural network. The adapters are introduced to perform feature transformation and dimensionality reduction on the time-domain and frequency-domain feature vectors. Optionally, the linear layer dimension of the two adapters is set to 128. t Representing F t The time-domain flight trajectory output vector after Adaptor1, O f Representing F f Frequency domain flight trajectory output vector after passing through Adaptor2;

[0117] 7.4) The time-domain flight trajectory output vector O obtained in step 7.3) t and frequency domain flight trajectory output vector O f The data is then concatenated and fed into the linear classification layer of neural network N to classify the flight trajectory signal.

[0118]

[0119] Where: represents vector concatenation operation; Linear represents a linear classification layer with an input dimension of 256 and an output dimension of 8, which is the total number of flight trajectory categories; This represents the type of flight trajectory predicted by the neural network.

[0120] 7.5) Calculate the prediction results for all data in the flight trajectory database according to steps 7.1)-7.4) and calculate the total classification error E. Optionally, the total classification error E can be calculated using the cross-entropy formula:

[0121]

[0122] 7.6) Backpropagate the training error E to the two-channel neural network and update the parameters θ of the two-channel neural network using stochastic gradient descent.

[0123] 7.7) Repeat steps 7.1)-7.6) until the total classification error no longer decreases. Optionally, an error threshold ε can be set. When the total classification error is below the threshold ε after 3 consecutive iterations, the training of the dual-channel neural network N can be considered complete.

[0124] 7.8) Input the trained flight trajectory recognition model, i.e., the dual-channel neural network N, into the flight trajectory recognition module.

[0125] The flight trajectory recognition module 5 in the host computer is used to recognize new flight trajectory signals using the flight trajectory recognition model input from the supervised training module, and is completed through the following process:

[0126] 8) The aircraft radiation source receiver monitors the flight trajectory signals of incoming aircraft in real time, and receives new flight trajectory signals s new Then input it into the flight trajectory recognition module;

[0127] 9) The flight trajectory recognition module will... new The signal is fed into a trained dual-channel neural network N for identification, resulting in the flight trajectory type of the new flight trajectory signal. In this example, the flight trajectory identification model will identify the type of the flight trajectory. new Recognized as type t6;

[0128] 10) The flight trajectory recognition module will display the recognition results and (s) new ,t6) is input as a new data sample into the flight trajectory database and stored. The stored new flight trajectory data will be used for the reinforcement training of the flight trajectory recognition model in the future.

[0129] The hardware of the host computer consists of the following components: a program memory for storing the implementation programs of each module; a data memory for storing flight trajectory data and parameters of the dual-channel neural network during model calculations; an arithmetic unit for executing programs and implementing corresponding functions; I / O components for receiving and transmitting data; and a display module for displaying the model training results and recognition results on the host computer.

[0130] The above embodiments are used to explain and illustrate the present invention, but not to limit the present invention. Any modifications and changes made to the present invention within the spirit and scope of the claims shall fall within the protection scope of the present invention.

Claims

1. A multi-target flight trajectory recognition device based on dual-channel pre-training, the recognition device comprising a flight trajectory database, an aircraft radiation source receiver, and a host computer; characterized in that: the flight trajectory database stores trajectory signal data of different types of aircraft, and the data stored in the database is used by the host computer; the host computer comprises a self-supervised pre-training module, a supervised training module, and a flight trajectory recognition module connected in sequence; the self-supervised pre-training module uses the signals in the flight trajectory database to perform self-supervised pre-training on the flight trajectory recognition model, and then uploads the flight trajectory recognition model after self-supervised pre-training to the supervised training module; the supervised training module uses the flight trajectory signals in the flight trajectory database to perform supervised training on the flight trajectory recognition model based on the self-supervised pre-trained flight trajectory recognition model, and then uploads the supervised training flight trajectory recognition model to the flight trajectory recognition module; finally, the aircraft radiation source receiver receives external aircraft radiation source signals, transmits them to the flight trajectory recognition module in the host computer, the flight trajectory recognition module displays the recognition results, tags the newly received aircraft flight trajectory signals with a type, and transmits them to the flight trajectory database for storage; The self-supervised pre-training module pre-trains the flight trajectory recognition model through three steps: signal transformation, signal masking, and signal reconstruction. The specific process is as follows: 1.1) Use Fast Fourier Transform to transform the original time-domain flight trajectory signals in the flight trajectory database. Transformed into frequency domain flight trajectory signal : Where FFT stands for Fast Fourier Transform; 1.2) In order to learn the intrinsic characteristics of flight trajectory signals through pre-training, for time-domain flight trajectory signals... and frequency domain flight trajectory signal All according to proportion Masking, that is, masking the time-domain flight trajectory signal and frequency domain flight trajectory signal All in proportion Random missing information: Mask represents the masking operation. Time-domain flight trajectory signal The signal behind the mask, Frequency domain flight trajectory signal The signal behind the mask; 1.3) Mask the time-domain flight trajectory signal and mask frequency domain flight trajectory signal Simultaneously, the signal is fed into a flight trajectory recognition model for signal reconstruction; the flight trajectory recognition model is a dual-channel neural network. It includes one time-domain channel and one frequency-domain channel; the time-domain channel receives the masked time-domain flight trajectory signal. Its self-supervised pre-training objective is to... Reconstructed into the original time-domain flight trajectory signal ; Frequency domain channel receive mask frequency domain flight trajectory signal Its self-supervised pre-training objective is to... Reconstructed into the original frequency domain flight trajectory signal This dual-channel neural network The self-supervised pre-training process is carried out in the following steps: 1.3.1) For all masked time-domain flight trajectory signals Reconstruction is performed, and the reconstruction error is calculated using the mean square error formula. : in For neural networks The reconstructed time-domain flight trajectory signal, For time-domain reconstruction error; 1.3.2) For all masked frequency domain flight trajectory signals Reconstruction is performed, and the reconstruction error is calculated using the mean square error formula. : in For neural networks The reconstructed frequency domain flight trajectory signal, For frequency domain reconstruction error; 1.3.3) Calculate the total reconstruction error The reconstruction error is used to backpropagate the two-channel neural network, and the parameters of the two-channel neural network are updated using stochastic gradient descent. ; 1.3.4) Repeat steps 1.3.1)-1.3.3) until the total reconstruction error no longer decreases, using the dual-channel neural network. Pre-training complete; 1.3.5) The self-supervised pre-training module will train the dual-channel neural network from step 1.3.4) after pre-training. and its network parameters Save the data and transfer it to the supervised training module; The supervised training module is based on a dual-channel neural network, which includes a time domain channel and a frequency domain channel.

2. The multi-target flight trajectory recognition device based on dual-channel pre-training as described in claim 1, characterized in that, The flight trajectory database is responsible for storing a large number of different types of flight trajectory signals: One of the parentheses represents a data point on an aircraft's electromagnetic radiation source. The time-domain signal representing the flight trajectory consists of multiple signal amplitudes sampled at a certain time step T. ,one Represents the amplitude of a sampling point. This represents the length of the time-domain signal of the flight trajectory, i.e., the total number of sampling points. represent Corresponding flight path type; This represents the total number of data points; the total number of flight trajectory categories is K.

3. The multi-target flight trajectory recognition device based on dual-channel pre-training as described in claim 1, characterized in that, The supervised training module trains the flight trajectory recognition model through three steps: signal transformation, signal classification, and parameter update. The specific process is as follows: 3.1) Use Fast Fourier Transform (FFT) to transform the original time-domain flight trajectory signals in the flight trajectory database. Transformed into frequency domain flight trajectory signal (Same as step 1.1), the transformed flight trajectory data is , For flight trajectory type labels; 3.2) For the dual-channel neural network fed in by the self-supervised pre-training module Supervised training shall be conducted in the following steps: 3.2.1) Transform the time-domain flight trajectory signal Input dual-channel neural network The time-domain channel extracts the time-domain features of the flight trajectory signal; the time-domain channel is calculated according to the following process. Extracting time-domain features from the data: in , , Represents the learnable parameter matrix; , , This represents an intermediate vector generated during the calculation process. Representing the Each time step Indicates flight trajectory signal No. The amplitude at each time step Dot product, Representing neural networks from The time-domain feature vector extracted from it; The Sigmoid function is represented by the following formula: 3.2.2) Frequency domain flight trajectory signal Input dual-channel neural network The frequency domain channel is used to extract the frequency domain features of the flight trajectory signal; the frequency domain channel is processed according to the following formula. Extracting frequency domain features from the data: in , , , , Represents the learnable parameter matrix, superscript Represents the transpose of a matrix; Representing neural networks from The frequency domain feature vector extracted from it; The Softmax function is represented by the following formula: 3.2.3) The time-domain feature vector obtained in step 3.2.1) And the frequency domain feature vector obtained in step 3.2.2). The data are fed into two adapters respectively, and feature transformation and dimensionality reduction are performed on the feature vectors of the two domains: in and Representing adapter 1 and adapter 2, both adapters are composed of a single-layer linear neural network; represent go through The output time-domain flight trajectory vector is then generated. represent go through The output vector of the frequency domain flight trajectory; 3.2.4) Output the time-domain flight trajectory vector obtained in step 3.2.3). and frequency domain flight trajectory output vector splicing, then feeding into the neural network The flight trajectory signal is classified in the linear classification layer: in Represents vector concatenation; Linear represents a linear classification layer; This represents the type of flight trajectory predicted by the neural network. 3.2.5) Calculate the prediction results for all data in the flight trajectory database according to steps 3.2.1)-3.2.4) and use the cross-entropy formula to calculate the total classification error. : 3.2.6) Based on training error Backpropagation is performed on the two-channel neural network, and the parameters of the two-channel neural network are updated using stochastic gradient descent. ; 3.2.7) Repeat steps 3.2.1)-3.2.6) until the total classification error no longer decreases, using a dual-channel neural network. Training complete; 3.2.8) The trained flight trajectory recognition model, i.e., the dual-channel neural network The data is then transmitted to the flight trajectory recognition module.

4. A multi-target flight trajectory recognition device based on dual-channel pre-training as described in claim 1, characterized in that, The aircraft radiation source receiver is responsible for receiving new flight trajectory signals and transmitting them to the flight trajectory recognition module. The flight trajectory recognition module receives the trained flight trajectory recognition model from the supervised training module and uses the model to recognize the new flight trajectory signals. The specific process is as follows: 4.1) The aircraft radiation source receiver monitors incoming aircraft radiation source signals in real time and receives new radiation source signals. It is then transmitted to the flight trajectory recognition module; 4.2) The flight trajectory recognition module will Feed into the trained dual-channel neural network The type of the new flight trajectory signal is identified during the process. ; 4.3) The flight trajectory recognition module will display the recognition results and... As a new data sample, it is input into the flight trajectory database and stored. The stored new flight trajectory data will be used for the reinforcement training of the flight trajectory recognition model in the future.

Images