Multi-source track association method based on deep learning

By using the GRU-Transformer model and Mahayana distance in track association, the problem of mis-association and leakage association in complex marine environments is solved, and more efficient and accurate track association is achieved, suitable for multi-source tracks Related tasks.

CN119004207APending Publication Date: 2024-11-22HARBIN UNIV OF SCI & TECH
View PDF 0 Cites 3 Cited by

Patent Information

Application Number
CN202411089353.3
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-08
Publication Date
2024-11-22

AI Technical Summary

Technical Problem

Traditional track correlation algorithms perform poorly when dealing with complex and dynamic maritime environments, especially in scenarios with large data volume and high target density, and are prone to mis-association and missed associations. The parameter settings are complex and rely on a large number of simulations and manual settings.

Method used

Using the multi-source track correlation method based on GRU-Transformer, normalized preprocessing of track data, GRU extracts features, combines Transformer model for correlation analysis, and realizes correlation classification through Marshall distance and classifier, and optimizes Network structure to improve generalization ability and reduce overfitting.

Benefits of technology

It improves the accuracy and efficiency of track associations, reduces the occurrence of mis-association and leaky associations, reduces the demand for human resources, and is suitable for multi-source track association tasks in complex maritime environments.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119004207A_ABST
    Figure CN119004207A_ABST
Patent Text Reader

Abstract

The invention discloses a multi-source track association method based on deep learning, and belongs to the field of track association algorithms. According to multi-source track association, track information obtained from different sensors or data sources is associated to the same target through an association algorithm; the deep learning method is a data driving method based on a neural network and is applied to a scene with high track density. The invention discloses a multi-source track association method based on deep learning. The method comprises an input layer part used for preprocessing track data; the GRU-Transform network model part is used for extracting the characteristics of the track data on the time sequence from the input data signal; a Mahalanobis distance correlation calculation section for taking into account the correlation between data features, suitable for data having a plurality of features; in the sigmod classification output part, classification output is carried out by utilizing features; associating unknown track data with maneuvering targets one by one; performing contrast experiment to test the performance of the method; according to the invention, track association is realized, and the association accuracy is improved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The invention relates to a multi-source track association method based on deep learning. Background Art

[0002] In the context of globalization and modernization, the increasing complexity of maritime transport activities has brought unprecedented challenges to the monitoring and management of ships. In the civilian field, international trade is frequently carried out, the volume of maritime transport continues to rise, and various types of ships such as merchant ships, tankers, cargo ships and fishing boats are constantly shuttling on the sea. The high-density maritime traffic flow has brought tremendous pressure to the maritime management department, especially in important waterways, ports and other sea areas. Track association obtains relevant information such as the track of the ship by matching the track of the ship, ensuring the safety of the ship's navigation and preventing maritime accidents.

[0003] In the military field, foreign warships have frequently illegally entered my country's territorial waters, seriously threatening national security and maritime navigation order. In responding to current challenges, information warfare and electronic warfare play a key role, and track correlation technology is an important part of it. Track correlation technology correlates relevant information of enemy ships from multiple sources, and can identify the movements and intentions of enemy ships, so as to adopt corresponding tactical strategies.

[0004] Using only a single source to observe the target track may result in the target track being lost or an incomplete target track being observed. Therefore, it is necessary to combine the information observed by multiple sources for comprehensive analysis. Track fusion can eliminate the accidental errors and uncertainties of the detection information of a single sensor, making the target's motion trajectory clearer and more reliable. It can also improve the ability to judge and predict the target state and reduce misjudgment or omissions caused by errors. Track fusion technology can integrate the observation data of different sensors to provide more comprehensive and accurate target information. It can also enhance the ability to monitor and identify targets, accurately judge the identity, motion state and intention of the target, and provide a more reliable basis for subsequent decision-making.

[0005] Track association is an important prerequisite for track fusion. Track association technology uses radar, satellite, AIS (Automatic Identification System) and other sensor data to match and associate track data from different sensors. It can identify the track set of the same target on different sensors, thereby determining the target's true position and motion trajectory.

[0006] By effectively fusing and correlating multi-source signal data, the position, movement status and changing trend of ships can be accurately and real-timely grasped, providing reliable data support for maritime management departments; improving the accuracy and efficiency of ship monitoring, and effectively warning and responding to emergencies, reducing the accident rate; the application of track correlation technology provides strong technical support for maritime monitoring and management, and is an important means to ensure maritime safety and achieve effective management; track correlation and fusion technology obtains the real-time trajectory and related information of enemy ships, accurately identifies the movements and intentions of enemy ships, and improves the monitoring capabilities of enemy ships; in complex maritime environments, it effectively supports military decision-making and action deployment, and safeguards national maritime rights and interests and military security;

[0007] Most of the existing traditional track association algorithms are based on probability statistics, such as Kalman filtering, weighted method and correction method, which are used in environments with a low number of tracks. However, facing today's more complex and denser marine environment, traditional methods have defects such as large communication volume, large number of simulations and manual parameter setting. They may perform poorly when dealing with track association in complex and dynamic environments. Deep learning models are good at processing and analyzing massive data. In ship track association, a large amount of historical and real-time data needs to be processed. Deep learning methods can efficiently analyze these data and discover potential association patterns. Therefore, this paper uses deep learning technology to propose a multi-source track association method based on GRU-Transformer for the problem of ship track association. Summary of the invention

[0008] The purpose of the present invention is to solve the problem of how to discover the hidden connections between data from a large amount of track data collected by multiple sensors, so as to reflect the target state more quickly, more accurately and more comprehensively, and propose a multi-source track association method based on deep learning;

[0009] The above purpose is achieved through the following technical solutions:

[0010] Step 1: Preprocess the collected track data. In the actual track data observation work, different dimensions may be used in different scenarios, which leads to orders of magnitude differences between the characteristic attributes of the track values. In order to make the neural network model unaffected by the scene and coordinate changes and to improve the generalization ability of the model, the track data should be normalized and preprocessed.

[0011] First, the data of the two observation sources need to be normalized and preprocessed, as shown in the following formula;

[0012]

[0013] Randomly select one track from different track sets of the same scene to form a track pair, as shown in the following formula;

[0014]

[0015] in and Represent the i-th track in Ω1 and the j-th track in Ω2 respectively; the longest time series t is selected for each scene max As the sequence length, if the length is insufficient, it will be padded with 0;

[0016] Step 2: Use the track association network module based on GRU-Transformer to extract the features of the track and input the feature matrix into the correlation calculation module;

[0017] The GRU-Transformer network model is specifically designed as follows:

[0018] GRU introduces a gating mechanism to control the flow of information, thereby showing better performance when processing long sequence data; the basic structure of GRU includes reset gate, update gate and candidate state; Transformer model is good at capturing global dependencies in sequence data;

[0019] The reset gate controls the influence of the previous hidden state on the candidate hidden state at the current moment; the calculation formula of the reset gate is as follows:

[0020] r t =σ(W r ·[h t-1 ,x t ])

[0021] The update gate determines to what extent the hidden state at the previous moment and the candidate hidden state at the current moment are updated to the hidden state at the current moment; the calculation formula of the update gate is as follows

[0022] z t =σ(W z ·[h t-1 ,x t ])

[0023] The candidate hidden state is calculated by the current input and the previous hidden state (adjusted by the reset gate); the calculation formula of the candidate hidden state is as follows:

[0024]

[0025] The final hidden state is calculated by the hidden state of the previous moment and the candidate hidden state (adjusted by the update gate); the calculation formula of the hidden state at the current moment is as follows:

[0026]

[0027] Transformer mainly relies on the attention mechanism. Different from the traditional recurrent neural network (RNN) and convolutional neural network (CNN), it completely abandons the cyclic processing of sequence data and uses parallel computing to improve training efficiency and performance. It consists of an encoder and a decoder.

[0028] The encoder part is composed of multiple identical encoder layers stacked together; each encoder layer consists of two sublayers: a multi-head self-attention mechanism and a feedforward neural network;

[0029] Multi-head attention concatenates the attention results of h heads and obtains the final output through linear transformation as shown below;

[0030] MultiHead(Q,K,V)=Concat(head1,…,head h )W o

[0031] The feedforward neural network in each encoder layer consists of two linear transformations and a ReLU activation function as shown below

[0032] FFN(x)=max(0,xW1+b1)W2+b2

[0033] The decoder part is similar to the encoder, and is also composed of multiple identical decoder layers stacked together; each decoder layer consists of three sublayers: multi-head self-attention mechanism, encoder-decoder attention mechanism, and feedforward neural network;

[0034] Step 3: Calculate the correlation between the two tracks from different sensors;

[0035] The measurement method of the correlation calculation module is Mahalanobis distance, which considers the correlation between data features and is suitable for data with multiple features. In track association, there is correlation between different features of different tracks, so Mahalanobis distance is more applicable, and the neural network can map the original features to a more discriminative feature space through the embedding layer.

[0036] Flatten the feature matrix output by the feature extraction network into a one-dimensional vector, calculate its Mahalanobis distance, and output a one-dimensional vector C of the same dimension, as shown in the following formula;

[0037] C=|Flatten(s1)-Flatten(s2)|

[0038] Step 4: Use the classifier Sigmod function to classify the track association;

[0039] The classifier consists of two fully connected layers, with 64 and 1 neurons respectively. The activation function is the Sigmoid function, and the output is 0 or 1, 1 indicates association, and 0 indicates no association. The output range of the Sigmoid function is (0,1), which is suitable for binary classification tasks, and the derivative form is simple, so that the gradient can be transferred relatively stably during the back propagation process, as shown in the following formula;

[0040]

[0041] Step 5: Conducting a track correlation performance comparison experiment on the invention;

[0042] A multi-source track association method based on deep learning was tested for performance. The representative PDA algorithm in the field of data association, the LSTM algorithm in the field of artificial intelligence that is good at processing time series, the nearest neighbor association method NN, and the AE-3D-CNN algorithm were selected as comparison algorithms, and the calculated track association accuracy was compared.

[0043] Furthermore, the track association method also includes the step of optimizing the network structure model, specifically:

[0044] (1) Transformer model optimization method:

[0045] The feature extraction module is composed of a Transformer model with the same structure, and the encoder has the same structure as the original Transformer encoder. In the track association task, the model receives the complete track sequence at one time and generates a complete feature matrix, which does not involve the order of the step-by-step output of the sequence in the Seq2Seq task. Therefore, the Masked Multi-head Attention layer (4) is removed from the decoder.

[0046] The vector of one track is used as the Q of the attention mechanism, and the vector of the other track is used as the K and V of the attention mechanism. For the Q vector of one track and all the K vectors of the other track, the attention coefficient between them is calculated; the calculation process is as follows:

[0047]

[0048] After attention calculation, a new feature vector is obtained, which highlights the information with high track correlation and reduces the influence of the information with low correlation on the judgment. The decoder decodes the feature vector and restores the time series code;

[0049] (2) Optimization method for overfitting phenomenon:

[0050] The Dropout layer is an effective regularization technique used to prevent overfitting of neural network models; overfitting refers to the model performing well on the training data but performing poorly on the test data or new data; Dropout creates different "sub-networks" during each training process by randomly discarding neurons and their connections, thereby improving the generalization ability of the model;

[0051] In each training iteration, set Dropout some neurons, set the selection retention probability p to 0.5, and 50% of the neurons will be randomly discarded; generate mask m: for each neuron, generate an independent random number, if the random number is greater than p, the neuron is "discarded" (that is, its output is set to zero), as shown in the following formula;

[0052]

[0053] In this way, we can effectively prevent the model from overfitting and improve the generalization ability of the model;

[0054] The beneficial effects of the present invention are:

[0055] Multi-source track association associates track information obtained from different sensors or data sources to the same target through an association algorithm. The task of the association algorithm is to determine which tracks from different sensors belong to the same target, thereby forming a globally consistent track target. Track association technology is one of the important means of maritime supervision. When faced with scenes with large amounts of data and high target density, traditional track association algorithms have high algorithm complexity and large computational workload, and are prone to mis-association and missed association. In addition, parameter settings are relatively complex and largely rely on experience. With the rapid rise of artificial skill technology, new ideas have been brought to track association technology. With massive track data as a foundation, the advantages of deep learning technology can be effectively utilized to mine track data features, improve the accuracy of track association, and save a lot of human resources.

[0056] The present invention designs a multi-source track association method based on GRU-Transformer, applies deep learning to the research of multi-source track association, so as to solve the problems of traditional track association technology, such as large algorithm complexity and large amount of calculation when facing scenes with large data volume and high target density, and easy occurrence of problems such as false association and missed association; GRU uses a gating mechanism, which can effectively capture local dependencies in time series data; through update gate (Update Gate) and reset gate (Reset Gate), GRU can control the transmission and retention of information; Transformer model is good at capturing global dependencies in sequence data; through self-attention mechanism, Transformer can effectively process long-distance dependencies and identify complex relationships between different time points in track data; the network is made applicable to track data of different sensors, so as to achieve the purpose of improving the accuracy of multi-source track association. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] Figure 1 is a flow chart of the method involved in the present invention;

[0058] Figure 2 It is a track association algorithm framework based on GRU-Transformer constructed by the present invention;

[0059] Figure 3 is the track association accuracy of different algorithms involved in the present invention;

[0060] Figure 4 It is a diagram of an improved Transofrmer network model involved in the present invention;

[0061] Figure 5 This is the change of neurons before and after the Dropout method involved in the present invention. DETAILED DESCRIPTION

[0062] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, but not all of the embodiments; based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention;

[0063] Preferred embodiments of the present invention:

[0064] See also Figure 1-Figure 5 , the present invention provides a technical solution:

[0065] A multi-source track association method based on deep learning, the core of the method consists of five modules: a track data preprocessing module to complete missing points and remove abnormal points; a track association network module of GRU-Transformer is used to extract track features and increase the correlation of the same target track through an attention mechanism; a correlation calculation module is used to calculate the correlation of multi-source track feature matrices; a classifier module consists of two fully connected layers for outputting simulation results; by pairing multi-source tracks in pairs and determining the correlation of track pairs, the traditional track association problem is converted into a binary classification problem of time series, the multi-source track association task is realized, and a track association performance comparison experiment is conducted on the invention. The method flow is as follows: Figure 1 , the method framework is as follows Figure 2 ; The described method is achieved by the following steps:

[0066] Step 1: Preprocessing of track association data. The multi-source track association problem can be decomposed into multiple two-source track association problems for separate processing. Therefore, the two-source track association problem is mainly taken as an example. The single track sequence in the same scenario is defined as Where i = 1, 2; the track sequence observed by source 1 is expressed as The track sequence observed by source 2 is expressed as The track sequences collected by two types of observation sources in the same scene are divided into two sets Ω1 and Ω2. The i-th track in Ω1 is expressed as The jth track in Ω2 is expressed as The jth track point in the i-th track in the track set is expressed as Where t, x, y, v and c represent time, longitude, latitude, speed and heading respectively; a track pair is randomly selected from different track sets of the same scene, as shown in the following formula;

[0067]

[0068] in and Represent the i-th track in Ω1 and the j-th track in Ω2 respectively; the longest time series t is selected for each scene max As the sequence length, if the length is insufficient, it will be padded with 0; the composed track pairs are divided into two categories, namely, associated group and non-associated group. The track pair combination of two track sequences of the same target is an associated group, and the data label is set to 1; the track combination of two track sequences of different targets is a non-associated group, and the data label is set to 0; before the data is input into the neural network, it is necessary to first perform normalization preprocessing on the data of the two observation sources, as shown in the following formula;

[0069]

[0070] Step 2: Use the track association network module based on GRU-Transformer to extract the features of the track and input the feature matrix into the correlation calculation module;

[0071] The GRU-Transformer network model is specifically designed as follows:

[0072] GRU uses a gating mechanism to effectively capture local dependencies in time series data. Through the update gate and reset gate, GRU can control the transmission and retention of information. The Transformer model is good at capturing global dependencies in sequence data.

[0073] GRU introduces a gating mechanism to control the flow of information, thereby showing better performance when processing long sequence data; the basic structure of GRU includes a reset gate, an update gate, and a candidate state;

[0074] The reset gate controls the influence of the previous hidden state on the candidate hidden state at the current moment; the calculation formula of the reset gate is as follows:

[0075] r t =σ(W r ·[h t-1 ,x t ])

[0076] The update gate determines to what extent the hidden state at the previous moment and the candidate hidden state at the current moment are updated to the hidden state at the current moment; the calculation formula of the update gate is as follows

[0077] z t =σ(W z ·[h t-1 ,x t ])

[0078] The candidate hidden state is calculated by the current input and the previous hidden state (adjusted by the reset gate); the calculation formula of the candidate hidden state is as follows:

[0079]

[0080] The final hidden state is calculated by the hidden state of the previous moment and the candidate hidden state (adjusted by the update gate); the calculation formula of the hidden state at the current moment is as follows:

[0081]

[0082] Transformer mainly relies on the attention mechanism. Unlike traditional recurrent neural networks (RNNs) and convolutional neural networks (CNNs), it completely abandons the cyclic processing of sequence data and uses parallel computing to improve training efficiency and performance.

[0083] Encoder

[0084] The encoder part is composed of multiple identical encoder layers stacked together; each encoder layer consists of two sublayers: a multi-head self-attention mechanism and a feedforward neural network;

[0085] Multi-head attention concatenates the attention results of h heads and obtains the final output through linear transformation as shown below:

[0086] MultiHead(Q,K,V)=Concat(head1,…,head h )W O

[0087] The feedforward neural network in each encoder layer consists of two linear transformations and a ReLU activation function as shown below:

[0088] FFN(x)=max(0,xW1+b1)W2+b2

[0089] Decoder

[0090] The decoder part is similar to the encoder, and is also composed of multiple identical decoder layers stacked together; each decoder layer consists of three sublayers: multi-head self-attention mechanism, encoder-decoder attention mechanism, and feedforward neural network;

[0091] Step 3: Calculate the correlation between the two tracks from different sensors;

[0092] The measurement method of the correlation calculation module is Mahalanobis distance, which considers the correlation between data features and is suitable for data with multiple features. In track association, there is a correlation between different features of different tracks, so Mahalanobis distance is more applicable, and the neural network can map the original features to a more discriminative feature space through the embedding layer. It helps to reduce the data dimension and filter out irrelevant features, thereby improving the calculation efficiency of Mahalanobis distance. Traditional Mahalanobis distance calculation methods are usually based on linear models, while neural networks can capture more complex nonlinear relationships. It provides more flexible modeling capabilities for processing complex ship data in actual scenarios, such as track, speed, and ship status. Introducing the Mahalanobis distance calculation method into the neural network can better solve the challenges of traditional calculation methods in processing the complexity of ship data with the help of the neural network's advantages such as adaptive learning, feature embedding, nonlinear modeling, end-to-end training, and generalization capabilities.

[0093] Flatten the feature matrix output by the feature extraction network into a one-dimensional vector, calculate its Mahalanobis distance, and output a one-dimensional vector C of the same dimension, as shown in the following formula;

[0094] C=|Flatten(s1)-Flatten(s2)|

[0095] Step 4: Use the classifier Sigmod function to classify the track association;

[0096] The classifier consists of two fully connected layers, with 64 and 1 neurons respectively. The activation function is the Sigmoid function, and the output is 0 or 1, 1 indicates association, and 0 indicates no association. The output range of the Sigmoid function is (0,1), which is suitable for binary classification tasks, and the derivative form is simple, so that the gradient can be transferred relatively stably during the back propagation process, as shown in the following formula;

[0097]

[0098] Step 5: Conducting a track correlation performance comparison experiment on the invention;

[0099] Furthermore, a multi-source track association method based on deep learning is tested for its performance. In order to verify the performance of the proposed method for track association, the average association accuracy P in all scenarios is calculated by simulation. CA , the average association accuracy in all scenarios is as follows Figure 3 The overall effect of the artificial intelligence method is better than the traditional PDA algorithm and NN algorithm. After 70 epochs, the association accuracy of the neural network is stable. The proposed MTAPT invention method has the most accurate association result among the five algorithms, with an association accuracy of 0.9277. The association results of other algorithms are: PDA method 0.8326, LTSM model 0.8931, AE-3D-CNN algorithm 0.9104;

[0100] Furthermore, the track association method also includes the step of optimizing the network structure model, specifically:

[0101] (1) Transformer model optimization method:

[0102] The feature extraction module is composed of the Transformer model with the same structure, and the encoder has the same structure as the original Transformer encoder. In the track association task, the model receives the complete track sequence at one time and generates a complete feature matrix, which does not involve the order of the step-by-step output of the sequence in the Seq2Seq task. Therefore, the MaskedMulti-head Attention layer is removed in the decoder part. Figure 4;

[0103] The vector of one track is used as the Q of the attention mechanism, and the vector of the other track is used as the K and V of the attention mechanism. For the Q vector of one track and all the K vectors of the other track, the attention coefficient between them is calculated; the calculation process is as follows:

[0104]

[0105] After attention calculation, a new feature vector is obtained, which highlights the information with high track correlation and reduces the influence of the information with low correlation on the judgment. The decoder decodes the feature vector and restores the time series code;

[0106] (2) Optimization method for overfitting phenomenon:

[0107] The Dropout layer is an effective regularization technique used to prevent overfitting of neural network models; overfitting refers to the model performing well on the training data but performing poorly on the test data or new data; Dropout creates different "sub-networks" during each training process by randomly discarding neurons and their connections, thereby improving the generalization ability of the model;

[0108] In each training iteration, set Dropout some neurons, set the selection retention probability p to 0.5, and 50% of the neurons will be randomly discarded; Generate mask m: For each neuron, generate an independent random number, if the random number is greater than p, the neuron is "discarded" (that is, its output is set to zero);

[0109]

[0110] In this way, we can effectively prevent the model from overfitting and improve the generalization ability of the model. Figure 5 ;

[0111] The embodiments of the present invention disclose preferred embodiments, but are not limited thereto. A person skilled in the art can easily understand the spirit of the present invention based on the above embodiments and make different extensions and changes. However, as long as they do not deviate from the spirit of the present invention, they are all within the protection scope of the present invention.

Claims

1. A multi-source track association method based on deep learning, characterized by: The method is implemented by the following steps: Step 1: Preprocess the collected track data. In the actual track data observation work, different dimensions may be used in different scenarios, which leads to orders of magnitude differences between the characteristic attributes of the track values. In order to make the neural network model unaffected by the changes in scenes and coordinates, and to improve the generalization ability of the model; Step 2: Use the track association network module based on GRU-Transformer to extract the features of the track and input the feature matrix into the correlation calculation module; Step 3: Calculate the correlation between two tracks from different sensors. In track correlation, there is a correlation between different features of different tracks. The Mahalanobis distance is used to consider the correlation between data features. Step 4: Use the classifier Sigmod function to classify track association, transform the traditional track association problem into a binary classification problem of time series, and realize the multi-source track association task; Step 5: Conducting a track correlation performance comparison experiment on the invention; Furthermore, the track association method also includes optimizing the model of the network structure.

2. The multi-source track association method based on deep learning according to claim 1, characterized in that: In the step 1, the step of preprocessing the collected track data is specifically as follows: In order to make the neural network model unaffected by scene and coordinate changes and to improve the generalization ability of the model, the track data must be normalized and preprocessed; First, the data of the two observation sources need to be normalized and preprocessed as follows: From different track sets of the same scene, one track pair is randomly selected to form the following formula: in and They represent the i-th track in Ω1 and the j-th track in Ω2 respectively. For each scene, the longest time series tmax is selected as the sequence length, and zeros are added if the length is insufficient.

3. The multi-source track association method based on deep learning according to claim 1, characterized in that: In the step 2, the steps of extracting the features of the track by using the track association network module based on GRU-Transformer and inputting the feature matrix into the correlation calculation module are specifically as follows: GRU controls the flow of information by introducing a gating mechanism. The basic structure of GRU includes a reset gate, an update gate, and a candidate state. The Transformer model captures global dependencies in sequence data. The reset gate controls the influence of the previous hidden state on the candidate hidden state at the current moment; the calculation formula of the reset gate is as follows: r t =σ(W r ·[h t-1 ,x t ]) The update gate determines to what extent the hidden state at the previous moment and the candidate hidden state at the current moment are updated to the hidden state at the current moment; the calculation formula of the update gate is as follows z t =σ(W z ·[h t-1 ,x t ]) The candidate hidden state is calculated by the current input and the previous hidden state (adjusted by the reset gate); the calculation formula of the candidate hidden state is as follows: The final hidden state is calculated by the hidden state of the previous moment and the candidate hidden state (adjusted by the update gate); the calculation formula of the hidden state at the current moment is as follows: Transformer mainly relies on the attention mechanism. Different from the traditional recurrent neural network (RNN) and convolutional neural network (CNN), it completely abandons the cyclic processing of sequence data and uses parallel computing to improve training efficiency and performance. It consists of an encoder and a decoder. The encoder part is composed of multiple identical encoder layers stacked together; each encoder layer consists of two sublayers: a multi-head self-attention mechanism and a feedforward neural network; Multi-head attention concatenates the attention results of h heads and obtains the final output through linear transformation as follows: MultiHead(Q,K,V)=Concat(head1,...,head h )W O The feedforward neural network in each encoder layer consists of two linear transformations and a ReLU activation function as follows; FFN(x)=max(0,xW1+b1)W2+b2 The decoder part is similar to the encoder, and is also composed of multiple identical decoder layers stacked together; each decoder layer consists of three sub-layers: multi-head self-attention mechanism, encoder-decoder attention mechanism, and feedforward neural network.

4. The multi-source track association method based on deep learning according to claim 1, characterized in that: In the step 3, the step of calculating the correlation between the two tracks of different sensors is specifically as follows: The measurement method of the correlation calculation module is Mahalanobis distance, which considers the correlation between data features and is suitable for data with multiple features. In track association, there is correlation between different features of different tracks, and the neural network can map the original features to a more discriminative feature space through the embedding layer. The feature matrix output by the feature extraction network is flattened into a one-dimensional vector, its Mahalanobis distance is calculated, and the one-dimensional vector C of the same dimension is output as follows. C=||Flatten(s1)-Flatten(s2)|| 5. The multi-source track association method based on deep learning according to claim 1, characterized in that: In the step 4, the step of using the classifier Sigmod function to classify the track association is specifically as follows: The classifier consists of two fully connected layers with 64 and 1 neurons respectively. The activation function is the Sigmoid function, and the output result is 0 or 1, 1 indicates association, and 0 indicates no association. The output range of the Sigmoid function is (0,1), which is suitable for binary classification tasks, and the derivative form is simple. During the back propagation process, the gradient can be transferred relatively stably as shown in the following formula.

6. The multi-source track association method based on deep learning according to claim 1, characterized in that: In the step 5, the steps of conducting a track correlation performance comparison experiment on the invention are specifically as follows: In order to verify the performance of the proposed method for track association, the simulation calculated the average association accuracy P in all scenarios. CA The overall effect of the artificial intelligence method is better than the traditional PDA algorithm and NN algorithm. After 70 epochs, the association accuracy of the neural network is stable. The proposed MTAPT invention method has the most accurate association result among the five algorithms, with an association accuracy of 0.9277. The association results of other algorithms are: PDA method 0.8326, LTSM model 0.8931, AE-3D-CNN algorithm 0.9104.

7. The multi-source track association method based on deep learning according to claim 1, characterized in that: The track association method also includes the step of optimizing the network structure model, specifically: (1) Transformer model optimization method: The feature extraction module is composed of a Transformer model with the same structure, and the encoder has the same structure as the original Transformer encoder. In the track association task, the model receives the complete track sequence at one time and generates a complete feature matrix, which does not involve the order of the step-by-step output of the sequence in the Seq2Seq task. Therefore, the MaskedMulti-head Attention layer is removed in the decoder part. The vector of one track is used as the Q of the attention mechanism, and the vector of the other track is used as the K and V of the attention mechanism. For the Q vector of one track and all the K vectors of the other track, the attention coefficient between them is calculated; the calculation process is as follows: After attention calculation, a new feature vector is obtained, which highlights the information with high track correlation and reduces the influence of the information with low correlation on the judgment. The decoder decodes the feature vector and restores the time series code; (2) Optimization method for overfitting phenomenon: The Dropout layer is an effective regularization technique used to prevent overfitting of neural network models; overfitting refers to the model performing smoothly on training data but performing poorly on test data or new data; Dropout creates different "sub-networks" during each training process by randomly discarding neurons and their connections, thereby improving the generalization ability of the model; In each training iteration, set Dropout some neurons, set the selection retention probability p to 0.5, and 50% of the neurons will be randomly discarded; generate a mask m: for each neuron, generate an independent random number, if the random number is greater than p, the neuron is "discarded" (that is, its output is set to zero) as shown below.

Citation Information

Cited By

  • Track correction method and device for cluster flying target, and medium

    CN119493094A

  • Multi-source target trajectory fusion method and system based on deep learning

    CN119808012A

  • Asynchronous track association method and device based on triple network, storage medium and electronic equipment

    CN120180371A