A feature fusion method applied to radar and AIS track association

Abstract

The application discloses a feature fusion method applied to radar and AIS track correlation, and comprises the following steps: pre-processing radar and AIS data; carrying out denoising treatment through a denoising auto-encoder model; constructing track and scene data; establishing a track feature extraction module by using a multilayer neural network; establishing a scene feature extraction module by using a three-dimensional convolutional neural network; fusing the extracted track and scene features in a channel dimension by using a Concat layer; and completing a track classification task by using a full connection layer and a Sigmoid activation function based on the fused features. The application combines track and scene features through deep learning, suppresses noise by using an auto-encoder, improves radar and AIS track correlation precision in multiple scenes, solves the problems of low correlation precision and poor scene adaptability of a traditional correlation method which only uses track features, and guarantees good correlation effect.

Description

Technical Field

[0001] This invention belongs to the field of radar and automatic identification system (AIS) track association technology, and relates to multi-source information fusion technology, specifically a feature fusion method applied to radar and AIS track association. Background Technology

[0002] Multi-source information fusion technology comprehensively utilizes data acquired from different sensors and integrates the features of each data source to extract sufficient information. It has wide applications in fields such as ship navigation and target detection.

[0003] Radar and AIS are the main types of equipment used in ship navigation. AIS is a satellite-based positioning system. It has high positioning accuracy and can periodically receive information such as shipment number, ship position, speed, and heading. Furthermore, AIS can operate in extreme marine environments. However, it can only receive information from cooperating vessels and its update speed is slow. Radar is a traditional device for maritime navigation and surveillance, with the advantage of fast information updates. However, it cannot provide static information such as target identity and destination, and radar's positioning accuracy is far lower than AIS. The fusion of radar and AIS output data can significantly improve the navigation accuracy of ships.

[0004] Track fusion is a hot topic in multi-source information fusion, and track association is a prerequisite for track fusion. The goal of ship track association is to determine the source of the collected tracks. Traditional track association methods are mainly based on statistics, fuzzy logic, and grey theory. Statistical methods utilize differences in target states to eliminate random uncertainty. This method uses probability thresholds for association judgment and requires many observation samples. Fuzzy logic-based methods use empirically designed membership functions to calculate the membership degree between tracks for judgment; the quality of the model depends on the design of the membership function. Grey theory-based methods use grey matrices and grey functions to regularize the data and are suitable for track association with small samples. Typically, the parameter settings of the above methods are complex and largely dependent on experience.

[0005] Machine learning can effectively solve complex trajectory association problems. These methods mainly include nearest neighbor methods, support vector machines (SVMs), and artificial neural networks. Zhao S. et al. proposed a nearest neighbor-based trajectory association method, which uses Euclidean distance, cosine similarity, and target trajectory as indicators, and sets a reasonable maximum relevance threshold based on the target's historical state. However, the performance of this method depends on feature engineering and threshold setting. Xu Y. et al. proposed an online discriminative learning method for feature combination in multi-target tracking. They utilize the structured output of support vector machines to describe the feature combination problem in data association. Constraints are used to distinguish between correct and distorted associations, making the association more reliable. However, they did not consider the applicability to different scenarios.

[0006] With the development of artificial intelligence, some scholars have also applied deep learning to trajectory association. Shi Z. et al. used a Restricted Long Short-Term Memory (LSTM) neural network for trajectory prediction. They used density-based spatial clustering and linear least squares techniques in data preprocessing. LSTM networks can extract spatiotemporal information; however, they are computationally intensive and not suitable for short time series. Xiong W. et al. designed a dual-contrast neural network for trajectory association. This network maps trajectory information in a high-dimensional space and selects the nearest neighbor vector as the associated trajectory. This method has strong scene adaptability, but its noise resistance is poor in dense track environments. Summary of the Invention

[0007] Purpose of the invention: In order to overcome the problem that traditional methods mostly only utilize track features (including position, speed and heading angle), resulting in low track association accuracy and poor adaptability to different scenarios, this invention provides a feature fusion method for radar and AIS track association. By integrating trajectory and scene features through deep learning, this method can associate radar and AIS output trajectories. This method has better scene adaptability and higher association accuracy.

[0008] Technical Solution: To achieve the above objectives, this invention provides a feature fusion method for correlating radar and AIS track data, comprising the following steps:

[0009] S1: Preprocess the acquired radar and AIS data, fill missing values ​​with cubic spline interpolation, and align the radar and AIS data in time.

[0010] S2: Denoise the data preprocessed in step S1 by establishing a good denoising autoencoder model;

[0011] S3: Construct track and scene data, take the statistical distance between any two tracks in position, speed and heading as track data, and use the Euclidean distance between all tracks in position to reconstruct scene data;

[0012] S4: Establish a track feature extraction module using a multi-layer neural network, and extract track features from track data through the track feature extraction module;

[0013] S5: Establish a scene feature extraction module using a 3D convolutional neural network, and extract scene features from the scene data through the scene feature extraction module;

[0014] S6: Use the Concat layer to fuse the extracted track and scene features in the channel dimension;

[0015] S7: Based on the fused features, the track classification task is completed through a fully connected layer and a Sigmoid activation function.

[0016] Furthermore, the data preprocessing method in step S1 is as follows: the original radar and AIS data are four-dimensional: X coordinate, Y coordinate, time and data source. Missing values ​​are processed by cubic spline interpolation and the original data is normalized to the range of 0 to 1. The corresponding data in the Y coordinate dimension is placed after the data in the X coordinate dimension. After preprocessing, the data becomes three-dimensional: coordinate, source and time dimensions, where the data in the X and Y coordinates are integrated into one dimension.

[0017] Furthermore, the method for establishing the denoising autoencoder model in step S2 is as follows:

[0018] For noisy input data X, the encoder (denoted by E) learns as many basic features of the input data as possible and stores them in a hidden vector Z; the decoder (denoted by D) reconstructs the hidden vector Z, completing the denoising autoencoder process by minimizing the input and output reconstruction errors to recover the data; the hidden vector Z is obtained by the following formula:

[0019] Z = f(X) = s f (W1X+b1) <1>

[0020] Where X = [x1, x2, ... x u ] is the input data, f(·) is the encoding function that maps X to the hidden representation Z, and the encoder is parameterized by θ1={W1,b1}, where W1 is the weight matrix and b1 is the bias vector of the input layer;

[0021] Reconstructed vector for:

[0022]

[0023] Among them, output data g(·) is the decoding function that maps the hidden representation Z to the output data; the decoder is parameterized by θ2={W2,b2}, where W2 is the weight matrix and b2 is the bias vector; the purpose of the autoencoder is to train parameters θ1 and θ2 to make X and As close as possible; the loss function for reconstruction is defined by the following formula:

[0024]

[0025] During the training of the autoencoder, the stochastic gradient descent (SGD) algorithm is used to update the parameters and minimize the reconstruction loss.

[0026] Furthermore, the method for constructing the track data in step S3 is specifically as follows:

[0027] For the i-th AIS track and the j-th radar track, the statistical distances between the two tracks in terms of position, speed, and heading are as follows:

[0028]

[0029]

[0030]

[0031] in, These are the variances of the position, speed, and heading of the AIS track, respectively. These are the variances of the radar track's position, speed, and heading, respectively.

[0032] Construct trajectory data D at time t t,m,n :

[0033] D t,m,n =[d s (i,j),d v (i,j),d c (i,j)] T <7> .

[0034] Furthermore, in step S3, radar and AIS track information are used to construct associated scene data. The specific construction method is as follows:

[0035] Add a time dimension to make the scene data structure resemble a video; sort the target track points of radar and AIS in the X and Y directions at time t according to the spatial location of the track in the sea area.

[0036] The sorting function of AIS is as follows:

[0037] s A =fA (i) <8>

[0038]

[0039] Where f A (·) represents the sorting function for AIS tracks. It is f A The inverse function of (·). If Then f A (i)>f A (j); if and Then f A (i)>f A (j);

[0040] The sequence number can be obtained from the batch number of the AIS track. Similarly, the sorting function f of the radar track can be obtained separately. R (·) and inverse function

[0041] At time t, the Euclidean distance between the radar's i-th orbit and the AIS's k-th orbit is as follows:

[0042]

[0043] And as scene D 1 The elements in the l-th row and k-th column, similarly, represent the Euclidean distance in the radar data. Euclidean distance in AIS data It can also be obtained through the following methods:

[0044]

[0045]

[0046] Construct scene data D:

[0047] D = [D 1 D 2 D 3 ] <13>

[0048] Among them, tensor D 1 The size is m×n×t, and the tensor D is... 2 The dimensions are m×m×t, and the tensor is D. 3 The size is n×n×t, and the scene data D is a four-dimensional tensor with channels, time, width, and height.

[0049] Furthermore, the method for establishing the track feature extraction module in step S4 is as follows:

[0050] The trajectory feature extraction module consists of three fully connected layers and one Dropout layer. The number of neurons in the three fully connected layers are 8, 16 and 36, respectively. During training, the Dropout layer temporarily deletes some neural network units with a certain probability. Due to random dropping, each batch can train different networks through the stochastic gradient descent algorithm, thus the network can alleviate the overfitting problem.

[0051] Furthermore, the method for establishing the scene feature extraction module in step S5 is as follows:

[0052] Convolutional Neural Networks (CNNs) are a type of feedforward neural network with a deep structure. 3D CNNs add a temporal dimension to 2D CNNs, further capturing temporal features. The expression for 3D convolution is:

[0053]

[0054] Where k,i,j∈{0,1,...,L} l+1}, L l+1 =(L l +2p-m) / s0+1;Z l and Z l+1 L represents the input and output of the (l+1)th convolutional layer, respectively; l+1 Z represents l+1 The dimensions are (assuming the depth, length, and width of the data are the same); Z(k,i,j) corresponds to a pixel in the three-dimensional data, H is the number of channels in the three-dimensional data, and m, s0, and p are the kernel size, stride, and padding number, respectively.

[0055] Furthermore, in step S5, three-dimensional convolutional and three-dimensional pooling layers are used to extract the basic features of the scene. The activation function of each layer is the ReLU function, and the output data of the module is flattened by the Flatten layer.

[0056] Furthermore, the method for fusing track and scene features in step S6 is as follows: the output data of the track feature extraction module and the output data of the scene feature extraction module are concatenated in the channel dimension through the Concat layer to obtain the fused track-scene features.

[0057] Furthermore, in step S7, the method for radar and AIS tracks to complete the classification task is as follows: the tracks are classified using a fully connected layer and a sigmoid activation function to determine whether any two tracks are associated targets or not.

[0058] Based on the above, the solution of the present invention can be summarized in the following three points:

[0059] (1) Data Construction: This invention models the correlation between radar and AIS tracks as a classification problem. The differences in position, velocity, and heading between radar and AIS tracks are constructed as track data. The Euclidean distances between the same and different tracks are constructed as scene data. The track and scene features are combined and input into the designed neural network to correlate the radar and AIS tracks.

[0060] (2) Noise Reduction: This invention uses an auto-encoder (AE) to suppress noise and interference. The AE module includes an encoder, a hidden layer, and a decoder. Noisy data is the input to the auto-encoder, and uncontaminated data is the automatically encoded label. The encoder encodes the noisy data into a hidden vector, and the decoder restores the hidden vector to the original data as much as possible, using the reconstruction error to design a loss function.

[0061] (3) Deep learning and feature fusion: This invention uses a fully connected layer to map trajectory data to a high-dimensional space to obtain trajectory features. A three-dimensional convolutional neural network (3D-CNN) is used to extract scene features from the constructed scene data, and a concat layer is used to fuse the trajectory and scene features. Finally, the scene adaptability is improved by fusing features through deep learning.

[0062] Beneficial Effects: Compared with existing technologies, this invention combines track and scene features through deep learning and utilizes an autoencoder to suppress noise, improving the accuracy of radar and AIS track association in multiple scenarios. It solves the problems of low association accuracy and poor scene adaptability of traditional association methods that only utilize track features (including position, velocity, and heading angle), ensuring good association results. Its specific advantages include:

[0063] 1. This invention proposes a method for constructing scene data, which uses a three-dimensional convolutional neural network for feature extraction, and can effectively obtain the spatial positional relationship between tracks within each scene.

[0064] 2. This invention proposes a noise reduction method based on an autoencoder, which can effectively reduce the impact of noise on prediction accuracy by training on noisy data to reconstruct the original trajectory.

[0065] 3. This invention proposes a method for fusing trajectory and scene features through deep learning, which can significantly improve the scene adaptability of the model. Attached Figure Description

[0066] Figure 1 This is the overall flowchart of radar and AIS track association provided by the present invention;

[0067] Figure 2 This is a scene data structure diagram provided by the present invention;

[0068] Figure 3 This is a diagram of the denoising module based on an autoencoder provided by the present invention;

[0069] Figure 4 This is a diagram of the scene feature extraction module based on a three-dimensional convolutional neural network provided by the present invention;

[0070] Figure 5 This is the training set loss map for all the methods provided in this invention;

[0071] Figure 6 This is a validation set accuracy graph of all methods provided by this invention;

[0072] Figure 7 This is a confusion matrix diagram of the AE-3D-CNN prediction results provided by the present invention;

[0073] Figure 8 This is a training set loss diagram of the ablation experiment provided by the present invention;

[0074] Figure 9 This is a graph showing the accuracy of the validation set for the ablation experiment provided by this invention.

[0075] Figure 10 This is the simulated flight path diagram provided by the present invention;

[0076] Figure 11 This is the denoised flight path map provided by the present invention. Detailed Implementation

[0077] The present invention will be further illustrated below with reference to the accompanying drawings and specific embodiments. It should be understood that these embodiments are for illustrative purposes only and are not intended to limit the scope of the invention. After reading this invention, any modifications of the invention in various equivalent forms by those skilled in the art will fall within the scope defined by the appended claims.

[0078] This invention provides a feature fusion method for correlating radar and AIS tracks, such as... Figure 1 As shown, it includes the following steps:

[0079] S1: Preprocessing of acquired radar and AIS data:

[0080] The raw radar and AIS data are four-dimensional: X-coordinate, Y-coordinate, time, and data source. Missing values ​​are processed using cubic spline interpolation, and the raw data is normalized to the 0-1 range. The corresponding data in the Y-coordinate dimension is placed after the data in the X-coordinate dimension. After preprocessing, the data becomes three-dimensional: coordinate, source, and time dimensions. The data in the X and Y coordinates are integrated into one dimension. Finally, the processed data is input into the autoencoder module.

[0081] S2: Establish a denoising autoencoder model:

[0082] Figure 3 The structure of an autoencoder is shown. For noisy input data X, the encoder (denoted by E) learns as many basic features of the input data as possible and stores them in a hidden vector Z. The decoder (denoted by D) reconstructs the hidden vector Z, completing the denoising autoencoder process by minimizing the input and output reconstruction errors to recover the data. The hidden vector Z is obtained by the following formula:

[0083] Z = f(X) = s f (W1X+b1) <1>

[0084] Where X = [x1, x2, ... x u ] is the input data, f(·) is the encoding function that maps X to the hidden representation Z, and the encoder is parameterized by θ1={W1,b1}, where W1 is the weight matrix and b1 is the bias vector of the input layer;

[0085] Reconstructed vector for:

[0086]

[0087] Among them, output data g(·) is the decoding function that maps the hidden representation Z to the output data; the decoder is parameterized by θ2={W2,b2}, where W2 is the weight matrix and b2 is the bias vector; the purpose of the autoencoder is to train parameters θ1 and θ2 to make X and As close as possible; the loss function for reconstruction is defined by the following formula:

[0088]

[0089] During the training of the autoencoder, the stochastic gradient descent (SGD) algorithm is used to update the parameters and minimize the reconstruction loss.

[0090] The preprocessed data is denoised by establishing a good denoising autoencoder model.

[0091] S3: Constructing track data, specifically including the following steps:

[0092] For the i-th AIS track and the j-th radar track, the statistical distances between the two tracks in terms of position, speed, and heading are as follows:

[0093]

[0094]

[0095]

[0096] in, These are the variances of the position, speed, and heading of the AIS track, respectively. These are the variances of the radar track's position, speed, and heading, respectively.

[0097] Construct trajectory data D at time t t,m,n :

[0098] D t,m,n =[d s (i,j),d v (i,j),d c (i,j)] T <7>

[0099] Constructing scene data:

[0100] The method for constructing correlated scene data using radar and AIS track information is as follows:

[0101] Add a time dimension to make the scene data structure resemble a video; sort the target track points of radar and AIS in the X and Y directions at time t according to the spatial location of the track in the sea area.

[0102] The sorting function of AIS is as follows:

[0103] s A =f A (i) <8>

[0104]

[0105] Where f A (·) represents the sorting function for AIS tracks. It is f A The inverse function of (·). If Then f A (i)>f A (j); if and Then f A (i)>f A (j);

[0106] The sequence number can be obtained from the batch number of the AIS track. Similarly, the sorting function f of the radar track can be obtained separately. R (·) and inverse function

[0107] At time t, the Euclidean distance between the radar's i-th orbit and the AIS's k-th orbit is as follows:

[0108]

[0109] And as scene D 1 The elements in the l-th row and k-th column, similarly, represent the Euclidean distance in the radar data. Euclidean distance in AIS data It can also be obtained through the following methods:

[0110]

[0111]

[0112] Construct scene data D:

[0113] D = [D 1 D 2 D 3 ] <13>

[0114] Among them, tensor D 1 The size is m×n×t, and the tensor D is... 2 The dimensions are m×m×t, and the tensor is D. 3 The size is n×n×t; Figure 2 The scene data structure is shown. Scene data D is a four-dimensional tensor with channels, time, width, and height.

[0115] S4: Establish a trajectory feature extraction module using a multi-layer neural network:

[0116] The trajectory feature extraction module consists of three fully connected layers and one Dropout layer. The number of neurons in the three fully connected layers are 8, 16 and 36, respectively. During training, the Dropout layer temporarily deletes some neural network units with a certain probability. Due to random dropping, each batch can train different networks through the stochastic gradient descent algorithm, thus the network can alleviate the overfitting problem.

[0117] The trajectory features of the trajectory data are extracted using the trajectory feature extraction module.

[0118] S5: Establish a scene feature extraction module using a 3D convolutional neural network:

[0119] Convolutional Neural Networks (CNNs) are a type of feedforward neural network with a deep structure. 3D CNNs add a temporal dimension to 2D CNNs, further capturing temporal features. The expression for 3D convolution is:

[0120]

[0121] Where k,i,j∈{0,1,...,L} l+1}, L l+1 =(L l +2p-m) / s0+1;Z l and Z l+1 L represents the input and output of the (l+1)th convolutional layer, respectively; l+1 Z represents l+1 The dimensions are (assuming the depth, length, and width of the data are the same); Z(k,i,j) corresponds to a pixel in the three-dimensional data, H is the number of channels in the three-dimensional data, and m, s0, and p are the kernel size, stride, and padding number, respectively.

[0122] The scene features of the scene data are extracted using the scene feature extraction module. Figure 4 The process of 3D convolution is shown, using 3D convolution and 3D pooling layers to extract basic features of the scene. The activation function of each layer is the ReLU function, and the output data of the module is flattened by the Flatten layer.

[0123] S6: Utilize the Concat layer to fuse the extracted track and scene features along the channel dimension:

[0124] The output data of the track feature extraction module and the scene feature extraction module are concatenated in the channel dimension by the Concat layer to obtain the fused track-scene features.

[0125] S7: Based on the fused features, the track classification task is completed through a fully connected layer and a sigmoid activation function.

[0126] The tracks are classified using a fully connected layer and a sigmoid activation function to determine whether any two tracks are associated or unassociated targets.

[0127] Based on the above, in order to verify the actual effect of the present invention, this embodiment applies the above solution to specific examples and conducts multiple experiments, as follows:

[0128] 1. Experimental Scenario:

[0129] (1) Data Construction

[0130] Assume the number of targets in the observed sea area follows a Poisson distribution with an average of 40 targets. The target accelerations in the X and Y directions follow a Gaussian distribution with a standard deviation of 2 m / s². The initial position, initial velocity, and initial heading of the targets follow a uniform distribution, ranging from [-5 km, 5 km], [50 m / s, 100 m / s], and [0°, 360°]. The radar detection probability is 0.9, the probability of the target being an AIS cooperative vessel is 0.9, and the measurement period is 4 s. Although the radar and AIS measurement periods differ, the tracks can be fitted and time-aligned using methods such as trispline interpolation. Both radar and AIS track segments consist of 5 track points, and zero-mean Gaussian white noise exists in the X and Y direction position measurements. The standard deviations of the AIS measurement noise in the X and Y directions are 30 m and 30 m, respectively, while the standard deviations of the radar measurement noise in the X and Y directions are 140 m and 140 m, respectively. The radar also has a systematic error, uniformly distributed within the range of 0–200 m.

[0131] This invention conducted 1000 Monte Carlo experiments across 1000 different scenarios. In any scenario, radar and AIS trajectories were paired. If the radar and AIS detected the same target, it was marked as a related sample pair; otherwise, it was marked as an unrelated sample pair. The ratio of related to unrelated pairs was approximately 1:20. Therefore, this invention discarded some unrelated samples to alleviate the imbalance between positive and negative samples. Finally, 34,328 samples were obtained, including 15,949 positive samples and 18,379 negative samples. The dataset was randomly divided into training, validation, and test sets in a ratio of 6:2:2.

[0132] (2) Experimental setup

[0133] This invention uses the PyTorch framework to build the AE-3D-CNN model, and all experiments were conducted on a workstation equipped with an NVIDIA GTX3090 GPU and a 3.8GHz Intel i7-10700k CPU.

[0134] The autoencoder module has 10 input and 10 output nodes. The six fully connected layers of the autoencoder (including encoder and decoder) have 16, 32, 64, 32, 16, and 10 nodes respectively. The trajectory feature extraction module has an input size of 1×3 and an output size of 1×36. The scene feature extraction module has four 3D convolutional kernels with sizes of 3×3×3, 3×3×3, 1×3×3, and 1×4×4 respectively. It has 3 input channels. The four convolutional modules have 3, 1, 1, and 1 channels respectively. The last two fully connected layers have 8 and 1 nodes respectively. Except for the last layer which uses the Sigmoid function, the network's activation function is the ReLU linear correction function. The learning rate of this neural network is 9×10⁻⁶.-4 .

[0135] 2. Analysis of experimental results:

[0136] To verify the effectiveness of the AE-3D-CNN model, this embodiment compares three methods using only trajectory features and three methods using both trajectory and scene features. The methods using only trajectory features are Random Forest [26-27], Backpropagation (BP), and Bidirectional Short-Term Memory (Bi-LSTM). Methods using both trajectory and scene features include 3D-CNN-BiLSTM, 3D-CNN, and AE-3D-CNN. This embodiment uses the scikit-learn toolkit to implement the Random Forest method and implements the other models in the PyTorch framework. The Bi-LSTM network consists of two stacked layers. The input size is 3, the hidden layer size is 8, and the forget gate parameter is 0.5. Finally, the two fully connected layers are concatenated for classification. The Random Forest uses the default parameters from the scikit-learn toolkit.

[0137] Figure 5 The loss curves of the neural network-based method on the training set are shown. Figure 6 The accuracy curves of each method on the validation set are presented. The following conclusions can be drawn:

[0138] (1) The convergence loss curve of Bi-LSTM has a smaller epoch. However, its training loss is more significant than that of the AE-3D-CNN model in terms of convergence, but its validation accuracy is lower than that of the AE-3D-CNN model. This is because the Bi-LSTM network requires sufficient contextual information to extract trajectory features, and due to the lack of sufficient temporal information in short series, it does not provide a significant improvement.

[0139] (2) The BP network has a slow convergence speed and low prediction accuracy. In order to keep the training data consistent with the training labels, the model is overtrained, which can easily lead to overfitting.

[0140] (3) After adding scene features, the convergence speed of the 3D-CNN-BiLSTM and 3D-CNN models is improved. However, due to the influence of noise, the training loss is higher than that of the AE-3D-CNN model.

[0141] (4) The proposed AE-3D-CNN model has fast convergence speed, minimal loss, and the highest recognition accuracy. The AE-3D-CNN model makes full use of the autoencoder module to reduce the noise of the original trajectory. The Concat layer integrates trajectory and scene features, which can adapt to different scenarios.

[0142] To further validate the model's generalization ability, the random forest method used 5x cross-validation to obtain the average classification accuracy, while the neural network-based method used five random seeds and averaged their accuracy. We tested six methods on the test set, which was not used for training. Table 1 shows the average accuracy of different models. The AE-3D-CNN model achieved the highest test accuracy. The average accuracy of the AE-3D-CNN model was 2.92%–3.48% higher than that of the random forest, BP, and Bi-LSTM models. This is because the AE-3D-CNN model uses a 3D convolutional network to extract scene features, exhibiting strong scene adaptability. The average accuracy of the AE-3D-CNN model was 1.2%–1.65% higher than that of the 3D CNN BiLSTM and 3D-CNN models. This is because the autoencoder was used to denoise the raw data in the AE-3D-CNN model, reducing the impact of measurement errors. Figure 7 The confusion matrix of the AE-3D-CNN model shows high accuracy in trajectory association prediction. These performance characteristics demonstrate that the proposed AE-3D-CNN model possesses strong generalization ability.

[0143] Table 1

[0144]

[0145] 3. Hyperparameter sensitivity experiment:

[0146] In this experiment, the sensitivity of the proposed model to the learning rate was tested. The learning rate was 3×10⁻⁶. -4 9×10 -4 and 1.5×10 -3 Table 2 shows the impact of different learning rates on test accuracy, loss, and iteration time. As the learning rate increases, the convergence speed accelerates, but test accuracy decreases and the loss increases. When the learning rate is 9×10... -4 At this point, the model exhibits the best fitting ability, the loss reaches its minimum, and the convergence speed is relatively fast.

[0147] Table 2

[0148]

[0149] 4. Ablation experiment:

[0150] This invention further explores the impact of the scene feature extraction module and the autoencoder module on model performance. Figure 8 and Figure 9The loss and accuracy curves of the model after ablation are shown on the training and validation sets, respectively. It can be seen that the "track + scene feature" model with the added scene feature extraction module converges faster because the fused features can focus not only on the global features of the track but also capture local features. The results show that the "denoising module + track + scene feature" (AE-3D-CNN) model has the fastest convergence speed and the highest validation accuracy because the denoising module reduces the impact of noise on the model's predictions. In conclusion, the scene feature extraction module and the autoencoder module are indispensable.

[0151] Figure 10 and Figure 11 The images show the original, noisy track and the denoised track. The AIS track corresponds to source number 1, and the radar track corresponds to source number 2. The number after the "-" indicates the track number. If the numbers are the same, they originate from the same target. Otherwise, they originate from different targets. It can be seen that the denoised track is smoother, and the distance between the two radar and AIS tracks originating from the same target is shorter than the original track.

Claims

1. A feature fusion method applied to radar and AIS track association, characterized in that, Comprising the following steps: S1: pre-process the obtained radar and AIS data, fill in the missing values by cubic spline interpolation, and time-align the radar and AIS data; S2: denoising the pre-processed data in step S1 by the established denoising autoencoder model; S3: constructing the track and scene data, taking the statistical distance of AIS track and radar track in position, speed and heading as track data, and using the Euclidean distance of AIS track and radar track in position to reconstruct the scene data; S4: establishing a track feature extraction module using a multi-layer neural network, and extracting track features of the track data through the track feature extraction module; S5: establishing a scene feature extraction module using a three-dimensional convolutional neural network, and extracting scene features of the scene data through the scene feature extraction module; S6: fusing the extracted track and scene features in the channel dimension using a Concat layer; S7: based on the fused features, completing the track classification task through a fully connected layer and a Sigmoid activation function; The construction method of the track data in step S3 is as follows: For the i-th AIS track and the j-th radar track, the statistical distance of the two tracks in position, speed and heading is as follows: wherein, are the variances of the position, velocity and heading of the AIS track, respectively; are the variances of the position, velocity and heading of the radar track, respectively; Trajectory data D at construction time t t,m,n : D t,m,n = [d s (i,j),d v (i,j),d c (i,j)] T <7>; In step S3, the track information of radar and AIS is used to construct the associated scene data, and the construction method is as follows: Add time dimension to make the scene data structure like a video; according to the spatial position of the track in the sea area, sort the target track points of radar and AIS in X and Y directions at t time respectively; The sorting function of AIS is as follows: s A = f A (i) <8> where f A (·) is a ranking function of AIS tracks, is the inverse function of f A (·); if then f A (i) > f A (j); if and then f A (i) > f A (j); The sorting number of AIS track is obtained by the track batch number of AIS track, and the sorting function f of radar track is obtained respectively R (·) and inverse function The Euclidean distance between the l-th radar track and the k-th AIS track at t time is as follows: and the as scenario D 1 the first row and the k-th column element of the matrix, likewise, the Euclidean distance 2 of scenario D and the Euclidean distance 3 of scenario D is obtained by Construct the scene data D: D = [D 1 ,D 2 ,D 3 ] <13> where D 1 has a tensor size of m x n x t, D 2 has a tensor size of m x m x t, and D 3 has a tensor size of n x n x t, and the scene data D is a four-dimensional tensor with channels, time, width, and height.

2. The feature fusion method for radar and AIS track association according to claim 1, characterized in that, The data preprocessing method in step S1 is as follows: the dimensions of the original radar and AIS data are four: X coordinate, Y coordinate, time and data source, the missing values are processed by cubic spline interpolation, and the original data is normalized to the interval of 0-1, the corresponding data in Y coordinate dimension is placed behind the data in X coordinate dimension, after preprocessing, the data becomes three-dimensional: coordinate, data source and time dimension, wherein the data in X and Y coordinates is integrated into one dimension.

3. The feature fusion method for radar and AIS track association according to claim 1, characterized in that, The establishment method of the denoising autoencoder model in step S2 is as follows: For noisy input data X, the encoder learns the basic features of the input data and stores them in the hidden vector Z; the decoder reconstructs the hidden vector Z, which completes the denoising autoencoder process by minimizing the input and output reconstruction error to recover the data; the hidden vector Z is obtained by the following formula: Z = f(X) = s f (W1X + b1) <1> where X = [x1, x2,... x u ] is the input data, f(·) is an encoding function that maps X to a hidden representation Z, and the encoder is parameterized by θ1= {W1, b1}, where W1is a weight matrix and b1is a bias vector for the input layer. Reconstructed vector is: where the output data g(·) is a decoding function that maps the hidden representation Z to the output data; the decoder is parameterized by θ2= {W2, b2}, where W2is a weight matrix and b2is a bias vector; the goal of the autoencoder is to make X and as close as possible; the loss function for the reconstruction is defined by the following equation: In the training process of the autoencoder, the stochastic gradient descent (SGD) algorithm is used to update the parameters θ1 and θ2 to minimize the loss function of reconstruction.

4. The feature fusion method for radar and AIS track association according to claim 1, characterized in that, The establishment method of the track feature extraction module in step S4 is as follows: The track feature extraction module includes three fully connected layers and a Dropout layer, the number of neurons in the three fully connected layers is 8, 16 and 36 respectively, and in the training process, the Dropout layer temporarily deletes some neural network units according to the probability, and due to random dropout, each small batch can train different networks through the stochastic gradient descent algorithm, so that the network can alleviate the problem of overfitting.

5. The feature fusion method for radar and AIS track association according to claim 1, characterized in that, The method for establishing the scene feature extraction module in the step S5 is specifically as follows: The convolutional neural network (CNN) is a kind of feedforward neural network with deep structure, and the three-dimensional CNN adds a time dimension to the two-dimensional CNN, which can further capture the time feature, and the expression of the three-dimensional convolution is as follows: where k, i, j ∈ {0, 1,..., L l+1}, L l+1 = (L l + 2p - m) / s0 + 1; Z l and Z l+1 represent the input and output of the l+1th convolutional layer, respectively; L l+1 represents the dimension of Z l+1 ; Z(k, i, j) corresponds to a pixel of the three-dimensional data, H is the number of channels of the three-dimensional data, and m, s0, and p are the convolution kernel size, the convolution step, and the padding number, respectively.

6. The feature fusion method for radar and AIS track association according to claim 5, characterized in that, In the step S5, the basic features of the scene are extracted by using the three-dimensional convolution and the three-dimensional pooling layer, the activation function of each layer is the Relu function, and the output data of the module is flattened by the Flatten layer.

7. The feature fusion method for radar and AIS track association according to claim 1, characterized in that, In the step S6, the fusion method of the track and scene features is as follows: the output data of the track feature extraction module and the output data of the scene feature extraction module are spliced in the channel dimension by the Concat layer to obtain the fused track-scene features.

8. The feature fusion method for radar and AIS track association according to claim 1, characterized in that, In the step S7, the method for completing the classification task of the radar and AIS tracks is as follows: the tracks are classified by the fully connected layer and the sigmoid activation function, and it is judged whether any two tracks are associated targets or non-associated targets.

Images