Compact ground wave radar marine ship track prediction method and system

Abstract

The invention relates to the technical field of maritime ship target beyond visual range monitoring, in particular to a compact type ground wave radar maritime ship track prediction method and system. AIS track data are obtained and converted into distance, speed and azimuth angles under a radar coordinate system to serve as characteristic parameters; the multi-dimensional target features which are segmented into n historical time steps through preprocessing and a sliding window serve as input of a prediction model, and the model outputs a prediction result of a future single time step; in the prediction model, each encoder layer extracts time sequence features of different depths from input features, and modeling is performed on the time sequence features by using a plurality of different attention mechanisms; and the output of the last layer of encoder passes through a linear mapping layer, and the high-dimensional time sequence representation is converted into the output prediction characteristics of the future time step. The multi-head self-attention mechanism is utilized to dynamically distribute the weights of the multi-dimensional target features, the complex association among different features is captured, the relationship among different time steps is analyzed from multiple angles, and the anti-interference capability is enhanced.

Description

Technical Field

[0001] The invention relates to the technical field of over-the-horizon surveillance and monitoring of marine vessel targets, and in particular to a method and system for predicting marine vessel tracks using a compact ground wave radar. Background Art

[0002] The statements in this section merely provide background information related to the present invention and do not necessarily constitute prior art.

[0003] Compact ground wave radar is a type of high-frequency ground wave radar. It uses the characteristics of ground waves "crawl" along the sea surface to break through the horizon and conduct long-range detection of sea surface targets. This radar system detects and locates targets by emitting high-frequency radio waves and receiving signals reflected from the target. Compared with traditional large-scale array ground wave radars, compact ground wave radars have a smaller footprint and greater flexibility, but due to limited transmission power, the received signal is easily interfered by sea clutter, ground clutter and background noise, resulting in a low signal-to-noise ratio of the target echo and discontinuous target observation data, which increases the difficulty of target tracking. Generally, various signal processing methods are used to improve its target detection performance.

[0004] Track prediction, by accurately predicting the state of the target at the next moment, can provide a reliable search range for the subsequent point-track-track association process, thereby improving the accuracy of point-track-track association. When the target is temporarily lost, the target position can be inferred based on the prediction results, reducing the probability of track interruption and ensuring the continuity of target tracking.

[0005] Traditional track prediction methods for marine vessels are mainly based on kinematic models, such as uniform speed models, uniform acceleration models, and turning models. These methods usually rely only on the current motion state of the target or very little historical data, and it is difficult to use the historical track information of the target to capture long-term dependencies. Therefore, they have limitations in dealing with nonlinear tracks and dealing with complex target dynamic changes. In particular, when the target undergoes drastic maneuvers (such as sharp turns, acceleration, or deceleration), traditional methods cannot quickly adapt to the frequently changing target motion state, resulting in a significant increase in track prediction errors. Summary of the invention

[0006] In order to solve the technical problems existing in the above-mentioned background technology, the present invention provides a compact ground wave radar marine ship track prediction method and system, which can dynamically allocate the weights of multi-dimensional target features by using a multi-head self-attention mechanism, capture the complex associations between different features, and analyze the relationship between different time steps from multiple angles. Under complex observation conditions (such as the presence of sea clutter, ground clutter or noise interference, etc.), it can focus on key information in a targeted manner, weaken the influence of irrelevant or low-correlation features, and enhance anti-interference ability.

[0007] In order to achieve the above object, the present invention adopts the following technical solution:

[0008] A first aspect of the present invention provides a compact ground wave radar marine vessel track prediction method, comprising the following steps:

[0009] The ship track data recorded by the Automatic Identification System (AIS) in the radar observation area is obtained and converted into distance, speed and azimuth in the radar coordinate system as feature parameters. The feature parameters are used as input data after preprocessing.

[0010] Based on the sliding window method, the continuous time series in the input data is divided into several samples, each of which contains multi-dimensional target features of n historical time steps, which are used as the input of the prediction model;

[0011] After the prediction model is trained, it outputs the prediction results of a single time step in the future and restores them to the distance, speed and azimuth corresponding to the actual physical quantities;

[0012] Among them, the prediction model has N layers of stacked encoders, each encoder layer extracts time series features of different depths from the input features; each encoder layer uses several different attention mechanisms to model the time series features; each encoder layer contains a feedforward neural network, which is used to perform nonlinear transformation on the features and further extract deep time series information; the output of the last layer of encoder is converted into the prediction features of the future time steps of the output through a linear mapping layer.

[0013] Furthermore, AIS track data is used as a data source for model training and testing, and the track data at least includes the longitude, latitude, speed and heading of the target ship.

[0014] Furthermore, during the training of the prediction model, the longitude, latitude and speed are converted into distance, azimuth and Doppler speed in the radar coordinate system as feature parameters and normalized.

[0015] Furthermore, based on the sliding window method, the continuous time series in the input data is divided into several samples. Specifically, each track data is regarded as a continuous time series, and the sliding window method is used to slide the data sequence according to the set window length and step size. Each slide generates a new set of data, and gradually generates training and test samples.

[0016] Furthermore, the prediction model is shown as follows:

[0017]

[0018]

[0019] in, represents the feature vector of the predicted future time step t, Θ represents the model parameters, and X t-n:t-1 Represents the observation data matrix of n historical time steps, x t-1 represents the feature vector of the t-1th time step, and F represents the feature dimension of each time step.

[0020] Furthermore, the prediction model is a DF-Transformer model, the input of the model is preprocessed multi-time step historical track data, and the output is a single time step target prediction state.

[0021] Furthermore, the data input to the prediction model is sequentially subjected to input embedding, position encoding, multi-layer encoder and linear mapping layer to obtain output data; wherein each layer of encoder includes a multi-head attention layer and a feedforward neural network connected in sequence, and both the multi-head attention layer and the feedforward neural network are connected to corresponding normalization layers.

[0022] Furthermore, during the training of the prediction model, the global dependency between time steps is captured by calculating the attention weights of each time step in the input sequence with respect to other time steps; each attention head models the relationship between sequence features from a different perspective, and a comprehensive representation is obtained by concatenating the results of multiple heads.

[0023] Furthermore, during the training of the prediction model, supervised training is performed using historical track samples and corresponding target prediction values; based on the mean square error loss function, the model parameters are gradually optimized through the gradient descent method.

[0024] A second aspect of the present invention provides a compact ground wave radar marine vessel track prediction system, comprising:

[0025] The data acquisition preprocessing module is configured to: obtain the AIS track data within the radar observation area, convert it into the distance, speed and azimuth in the radar coordinate system as characteristic parameters, and use the characteristic parameters as input data after preprocessing;

[0026] The sliding window module is configured to: split the continuous time series in the input data into several samples based on the sliding window method, each sample contains multi-dimensional target features of n historical time steps, and use it as the input of the prediction model;

[0027] The track prediction module is configured as follows: after the prediction model is trained, the prediction result of a single time step in the future is output and restored to the distance, speed and azimuth corresponding to the actual physical quantity;

[0028] Among them, the prediction model has N layers of stacked encoders, each encoder layer extracts time series features of different depths from the input features; each encoder layer uses several different attention mechanisms to model the time series features; each encoder layer contains a feedforward neural network, which is used to perform nonlinear transformation on the features and further extract deep time series information; the output of the last layer of encoder is converted into the prediction features of the future time steps of the output through a linear mapping layer.

[0029] Compared with the prior art, one or more of the above technical solutions have the following beneficial effects:

[0030] 1. Capable of capturing long-term dependencies and improving adaptability to complex targets. The prediction model effectively captures long-term dependencies in target track data through a multi-layer encoder structure. The model is based on historical track information and makes predictions based on the target's motion patterns and changing trends. It has good adaptability in complex, nonlinear motion modes. Compared with traditional deep neural networks, this model is not restricted by the time step sequence during the reasoning process, can quickly generate target prediction results, and is suitable for use in compact ground wave radar systems with limited computing resources.

[0031] 2. Use the multi-head self-attention mechanism to dynamically model feature relationships and improve anti-interference capabilities. The multi-head self-attention mechanism in the prediction model can dynamically allocate weights of multi-dimensional target features, capture the complex associations between different features, and analyze the relationships between different time steps from multiple perspectives. Under complex observation conditions (such as the presence of sea clutter, ground clutter, or noise interference), the model can focus on key information in a targeted manner, weaken the impact of irrelevant or low-correlation features, and enhance anti-interference capabilities.

[0032] 3. The DF-Transformer model used in the prediction model omits the decoder module of the traditional Transformer model, and only uses the encoder part to model and process the input sequence, and maps the encoder output to the target prediction state through the linear mapping layer. The lightweight network structure not only retains the ability of the Transformer model to capture global dependencies, but also significantly reduces the number of model parameters and computational complexity, achieving a combination of model efficiency and prediction performance.

[0033] 4. The sliding window sampling method is used to divide the track data into time series segments of fixed length for model training and testing. Sliding window sampling can generate diverse samples from track data of different targets and different time periods, so that the model can fully learn the motion characteristics of different targets and dynamic laws in various scenarios. This design improves the generalization ability of the model, enabling it to adapt to complex application requirements of multiple targets and multiple scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] The accompanying drawings in the specification, which constitute a part of the present invention, are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute improper limitations on the present invention.

[0035] Figure 1 is a schematic diagram of a process of predicting a marine vessel track using a compact ground wave radar provided by one or more embodiments of the present invention;

[0036] Figure 2 Schematic diagram of a network architecture during track prediction provided by one or more embodiments of the present invention. DETAILED DESCRIPTION

[0037] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0038] It should be noted that the following detailed descriptions are exemplary and are intended to provide further explanation of the present invention. Unless otherwise specified, all technical and scientific terms used herein have the same meanings as those commonly understood by those skilled in the art to which the present invention belongs.

[0039] Terminology explanation:

[0040] LSTM, Long Short-Term Memory, is a special recurrent neural network (RNN) that aims to solve the problem of gradient vanishing or gradient exploding when traditional RNN processes long sequence data, so that the network can learn long-term dependencies.

[0041] The Transformer model is a neural network model based on the self-attention mechanism. The core is the self-attention mechanism, which allows the model to consider all positions in the sequence when calculating the representation of each element, so as to capture long-distance dependencies. This mechanism enables the Transformer model to calculate the association between any two positions in the sequence without considering their distance in the sequence, so it can better capture long-distance dependencies.

[0042] The DF-Transformer (Decoder-Free Transformer) model is a variant that removes the decoder part of the traditional Transformer model. In the traditional Transformer model, the encoder is responsible for mapping the input sequence to a hidden representation that contains the semantic information of the input sequence; and the decoder generates the target sequence based on the output of the encoder. However, in some specific tasks, such as text classification and sentiment analysis, it is not necessary to generate the target sequence, but only to classify or analyze the input sequence. In this case, the Decoder-Free Transformer model is particularly applicable.

[0043] As introduced in the background technology, the traditional track prediction method for marine vessels is mainly based on kinematic models. It is usually difficult to use the historical track information of the target to capture long-term dependencies, and there are limitations in dealing with nonlinear tracks and dealing with complex target dynamic changes. When the target undergoes drastic maneuvers (such as sharp turns, acceleration or deceleration), it is impossible to quickly adapt to the frequently changing target motion state, resulting in a significant increase in the track prediction error.

[0044] Long Short-Term Memory (LSTM) has certain advantages in capturing time series dependencies, but its sequential computing characteristics limit its parallel processing capabilities. It also has obvious deficiencies in processing complex feature relationships and dynamic weight allocation. In contrast, the Transformer model processes the entire sequence in parallel through a self-attention mechanism, without the need for recursion based on time steps, which improves computational efficiency and long-term dependency modeling capabilities. In addition, its multi-head self-attention mechanism can model the interaction relationship of multi-dimensional features in time series from different angles, such as the distance, speed, and azimuth of the target. Therefore, the Transformer model shows significant advantages in track prediction tasks in complex scenarios, providing strong technical support for improving the target tracking continuity of compact ground wave radars.

[0045] Embodiment 1:

[0046] A compact ground wave radar marine vessel track prediction method comprises the following steps:

[0047] The AIS track data of ship targets in the radar observation area are collected and processed, and converted into characteristic parameters (such as distance, Doppler velocity and azimuth) in the radar coordinate system as the original data source of the model. The data is processed using the normalization method to meet the model input requirements, and divided into training sets and test sets to ensure the effectiveness of subsequent model training and testing.

[0048] The sliding window technique is used to divide the continuous time series in the training set data into several samples. Each sample contains multidimensional target features of n historical time steps as the input of the DF-Transformer model and corresponds to an output target prediction state of the future time step.

[0049] A lightweight trajectory prediction model based on DF-Transformer is constructed, retaining only the encoder part, including the multi-head self-attention mechanism and feedforward neural network. The training set data is input into the model and trained using supervised learning. After the training is completed, a model that can predict the future state of the target is obtained.

[0050] The trained prediction model is verified using the test set, and the historical track data sequence is input to predict the future state of the target. The prediction result of the target in the future single time step is output, and the prediction result is restored to the actual physical value (target distance, speed, azimuth, etc.) through denormalization.

[0051] In this embodiment, the process of track prediction is as follows: Figure 1 As shown, specifically: obtain the AIS data set and preprocess it, divide the preprocessed data into a training set and a test set, train the model after sliding window sampling and model construction to obtain a trained prediction model, and test the model to obtain the required prediction results. For the target track being tracked by the compact ground wave radar, once the existence of the target is preliminarily verified, the trained prediction model can be used to predict the track.

[0052] (1) Obtain data and preprocess it.

[0053] This embodiment uses Automatic Identification System (AIS) data as the main data source for network training and testing. AIS can provide real-time dynamic information of ship targets, including multi-dimensional features such as longitude, latitude, speed, and heading. Compared with radar measurement data, AIS data has the advantages of high precision, high-frequency sampling, and comprehensive features, and can provide high-quality training samples and test benchmarks for the track prediction task of compact ground wave radar.

[0054] Collect AIS track data containing target longitude, latitude, speed, heading and other features in the radar observation area, and convert longitude, latitude and speed into distance, azimuth, Doppler speed and other features in the radar coordinate system. Apply the MinMaxScaler method (maximum-minimum normalization method) to normalize the data and scale each feature value to a uniform numerical range to meet the input requirements of the Transformer model while ensuring the stability and convergence of the model training.

[0055] (2) Sliding window sampling.

[0056] Each AIS track data is regarded as a continuous time series, and the sliding window method is used to slide the data sequence according to the specified window length with a fixed step size. Each slide will generate a new set of data, and gradually generate training and test samples. By controlling the window length and sliding step size, it is ensured that the input sequence can cover the historical track data of the target, and the window is gradually slid to generate prediction samples. In the specific implementation, the sliding window length is set to n time steps, the prediction window is set to 1 time step, and predictions are performed on all time steps except the first n points, so as to generate data samples covering the entire track.

[0057] (3) Model construction.

[0058] The track prediction task is formulated as a multivariate time series prediction problem, which predicts the state of the ship target at a future moment based on its historical state. The prediction model can be expressed as:

[0059]

[0060]

[0061] in, represents the feature vector of the predicted future time step t, Θ represents the model parameters, including multi-head self-attention, weights of feedforward neural network, etc.; X t-n:t-1 Represents the observation data matrix of n historical time steps, x t-1 represents the feature vector of the t-1th time step, and F represents the feature dimension of each time step (such as distance, Doppler velocity, azimuth, etc.).

[0062] The standard Transformer model usually includes an encoder and a decoder structure, which has a large number of parameters and high computational complexity, and is difficult to be directly applied to resource-constrained radar signal processing tasks. Therefore, this embodiment adopts a lightweight Transformer model: Decoder-Free Transformer, referred to as DF-Transformer. This model only retains the encoder part and omits the decoder module of the standard Transformer model. The input of the model is preprocessed multi-time step historical track data, and the output is a single time step target prediction state.

[0063] Traditional Transformer is usually used for text translation, including encoder and decoder, and the decoder is used to translate the output sequence word by word. The track prediction task of compact ground wave radar does not need to generate complex output sequences. It only needs to predict the state of the target at the next moment based on its n historical states. The decoder module is not necessary in this task, and the task focuses more on feature extraction using the encoder. Therefore, the lightweight design that omits the decoder can significantly reduce the number of parameters and computational complexity. Secondly, compact ground wave radar systems are usually deployed on small platforms or space-constrained environments with limited hardware resources (such as computing power and power consumption). At the same time, considering the real-time nature of the prediction, a smaller network is easier to deploy on site.

[0064] In addition, the compact ground wave radar can detect multiple characteristic parameters such as the distance, Doppler velocity and azimuth of the target at sea, and simultaneously predict the multi-dimensional features during the track prediction process, so as to better grasp the real-time motion state of the target. However, there are complex interactions between different parameters, which are difficult to describe with a unified formula. To this end, the DF-Transformer model of this embodiment utilizes a multi-head self-attention mechanism, and different attention heads model the characteristic relationship in the time series from different angles, dynamically adjust the weight of the track feature, and enhance the model's ability to capture target features in complex scenarios, and make the DF-Transformer model more suitable for track prediction of compact ground wave radar.

[0065] The network structure is as follows Figure 2 As shown in the figure, the input data is sequentially embedded, positionally encoded, multi-layered encoder and linearly mapped to obtain output data; wherein the multi-layered encoder is a multi-layered Transformer encoder model, and each layer of the encoder includes a multi-headed attention layer and a feedforward neural network connected in sequence, and the multi-headed attention layer and the feedforward neural network are connected with corresponding normalization layers.

[0066] The core of the multi-layer Transformer encoder model is the multi-head self-attention mechanism. For each time step in the input sequence, the attention weight with other time steps is calculated to capture the global dependency between time steps. Each attention head models the relationship between sequence features from different angles, and finally a comprehensive representation is obtained by splicing the results of multiple heads, which improves the ability to model feature interaction.

[0067] Combination Figure 2Each attention head of the multi-head self-attention independently models the relationship between sequence features from different angles, and finally splices the results of multiple attention heads together to obtain a comprehensive representation of each time step. The spliced ​​multi-head attention results will be further extracted from the output of the attention mechanism through the feedforward neural network in the encoder to obtain higher-level temporal features. After being processed by multiple layers of encoders, the output of the last layer of encoder is a high-dimensional temporal feature representation. In order to obtain prediction results (such as the distance, Doppler velocity, and azimuth of the target in the next time step), these high-dimensional features will be converted into output features through a linear mapping layer.

[0068] In the above process, the input time series will pass through multiple encoder layers, and then pass through the multi-head self-attention mechanism, feedforward neural network and normalization layer to extract deep feature representation. The output of the last layer of encoder will pass through a linear mapping layer and be reduced to the predicted value of the target state. The DF-Transformer model of this embodiment replaces the decoder part in the Transformer model with a linear mapping layer composed of a fully connected layer.

[0069] In this embodiment, the number of encoder layers N is 5, indicating that there are 5 stacked encoder layers, each of which is responsible for extracting time series features of different depths from the input features. The number of multi-head self-attention heads is 6, indicating that 6 different attention mechanisms are used in each encoder layer to model the time series features. The linear mapping layer converts the 6-dimensional input into a 5-dimensional output. Each encoder layer contains a feedforward neural network for nonlinear transformation of features and further extracting deep time series information. The output of the last layer of encoder passes through a linear mapping layer to convert the high-dimensional time series representation into the predicted features of the output future time steps.

[0070] In summary, the model design of this embodiment is suitable for small time series prediction tasks. It realizes efficient time series modeling capabilities through a lightweight network structure and is suitable for the real-time track prediction needs of compact ground wave radar.

[0071] (4) Model training.

[0072] The model is iteratively optimized by batch training AIS data samples, and supervised training is performed using historical track samples and corresponding target prediction values. The generated training data is input into the DF-Transformer model. During the training process, the model gradually optimizes the network parameters through the gradient descent method based on the mean square error loss function. In order to improve the convergence speed and stability of the training, the model parameters are initialized (such as Xavier initialization) to prevent overfitting. After the training is completed, the optimal model parameters are saved in the specified file for use in the test phase.

[0073] During training, the batch size is 64, the optimizer is Adam, and the initial learning rate is set to 10 -4 , and gradually increased to the initial learning rate through linear warm-up in the first 4000 steps of training. Afterwards, the learning rate was gradually reduced to 10 using the cosine annealing decay strategy. -8 , ensuring that the model can converge stably and avoid overfitting in the later stages of training.

[0074] (5) Model testing.

[0075] In the test phase, the track data to be tested is processed in the same sliding window manner and then input into the model. Based on the input target historical track data, the model can extract the dependency between features through the encoder, and then generate the target prediction value at the next moment through the linear mapping layer. The output prediction results are converted into actual physical quantities (target distance, Doppler velocity, azimuth, etc.) through denormalization and stored in the database for subsequent analysis and application.

[0076] This embodiment can capture long-term dependencies and improve adaptability to complex targets. The powerful time series modeling capability of the Transformer model is utilized to effectively capture long-term dependencies in target track data through a multi-layer encoder structure. The model is based on historical track information and makes predictions based on the movement patterns and change trends of the target. It has good adaptability in complex and nonlinear motion modes. Compared with traditional deep neural networks, this model is not restricted by the time step sequence during the reasoning process, can quickly generate target prediction results, and is suitable for use in radar systems with limited computing resources.

[0077] This embodiment uses a multi-head self-attention mechanism to dynamically model feature relationships and improve anti-interference capabilities. The multi-head self-attention mechanism in the Transformer model can dynamically assign weights to multi-dimensional target features, capture complex associations between different features, and analyze the relationships between different time steps from multiple perspectives. Under complex observation conditions (such as the presence of sea clutter, ground clutter, or noise interference), the model can focus on key information in a targeted manner, weaken the impact of irrelevant or low-correlation features, and enhance anti-interference capabilities.

[0078] This embodiment simplifies the Transformer network structure and reduces computational complexity. The DF-Transformer model omits the decoder module of the traditional Transformer model, and only uses the encoder part to model and process the input sequence, mapping the encoder output to the target prediction state through a linear mapping layer. This lightweight network structure not only retains the Transformer model's ability to capture global dependencies, but also significantly reduces the number of model parameters and computational complexity, achieving a combination of model efficiency and prediction performance.

[0079] The sliding window sampling in this embodiment can ensure data diversity and improve the generalization ability of the model. The sliding window sampling method is used to divide the AIS track data into time series segments of fixed length for model training and testing. Sliding window sampling can generate diverse samples from track data of different targets and different time periods, so that the model can fully learn the motion characteristics of different targets and dynamic laws in various scenarios. This design improves the generalization ability of the model, enabling it to adapt to complex application requirements of multiple targets and multiple scenarios.

[0080] Experimental results and analysis:

[0081] Since the motion of a ship target at sea can be approximated as uniform linear motion in a short period of time, traditional methods usually use a uniform linear motion model to predict the target state in a single time step in the future. This embodiment uses a uniform linear motion model as a comparison method to evaluate the performance of the proposed track prediction method based on the DF-Transformer model. In order to quantify the prediction performance, this embodiment uses the following evaluation indicators:

[0082] 1) Mean Absolute Error (MAE): reflects the average level of absolute error between the predicted value and the true value, and can measure the overall prediction deviation of the model.

[0083] 2) Mean Square Error (MSE): Calculates the square average of the error between the predicted value and the true value. It is more sensitive to larger prediction errors and can reflect the performance of the model when dealing with outliers or large deviations.

[0084] 3) Root Mean Square Error (RMSE): As the square root of MSE, it can more intuitively reflect the scale of the prediction error and help to comprehensively evaluate the prediction effect of the model.

[0085] During the experiment, a compact surface wave radar system (CORMS system) was deployed on a certain coast. Since the data quality and quantity of the radar measured data could not guarantee the full training of the model, the AIS (Automatic Identification System) data within the detection range of a certain day were selected for model training and testing. The data included the trajectories of 100 ship targets within 226 consecutive minutes, and the state of each target every minute was 1 sampling point. Among them, the AIS tracks of 70 targets were used for model training, and the AIS tracks of the remaining 30 targets were used for model testing. The length of the historical sequence n used for track prediction is 20, so the total number of training samples is (226-20)*70=14420. These AIS data contain multi-dimensional target features and reflect the target movement under different environmental conditions, providing rich test scenarios and data support for the experiment.

[0086] In the experiment, the above evaluation indicators were applied to the key characteristic parameters of the target, namely, distance, Doppler velocity and azimuth. The calculation results of the error indicators are shown in Table 1, which intuitively reflects the difference in prediction performance between the traditional method and the proposed track prediction method based on the DF-Transformer model in different feature dimensions.

[0087] Table 1 Comparison of prediction error indicators

[0088]

[0089] Experimental results show that the proposed method has obvious advantages in MAE, MSE and RMSE indicators, with smaller prediction error and higher prediction accuracy.

[0090] Therefore, this embodiment can efficiently and accurately complete the track prediction task of marine ship targets, and significantly improve the target tracking capability of the compact ground wave radar system in complex marine environments. Experimental verification shows that this method is superior to the traditional track prediction method based on the kinematic model in terms of error indicators (MAE, MSE, RMSE).

[0091] Embodiment 2:

[0092] A compact ground wave radar marine vessel track prediction system, comprising:

[0093] The data acquisition preprocessing module is configured to: obtain the AIS track data within the radar observation area, convert it into the distance, speed and azimuth in the radar coordinate system as characteristic parameters, and use the characteristic parameters as input data after preprocessing;

[0094] The sliding window module is configured to: split the continuous time series in the input data into several samples based on the sliding window method, each sample contains multi-dimensional target features of n historical time steps, and use it as the input of the prediction model;

[0095] The track prediction module is configured as follows: after the prediction model is trained, the prediction result of a single time step in the future is output and restored to the distance, speed and azimuth corresponding to the actual physical quantity;

[0096] Among them, the prediction model has N layers of stacked encoders, each encoder layer extracts time series features of different depths from the input features; each encoder layer uses several different attention mechanisms to model the time series features; each encoder layer contains a feedforward neural network, which is used to perform nonlinear transformation on the features and further extract deep time series information; the output of the last layer of encoder is converted into the prediction features of the future time steps of the output through a linear mapping layer.

[0097] The multi-head self-attention mechanism can dynamically allocate the weights of multi-dimensional target features, capture the complex associations between different features, and analyze the relationships between different time steps from multiple perspectives. Under complex observation conditions (such as the presence of sea clutter, ground clutter, or noise interference), it can focus on key information in a targeted manner, weaken the influence of irrelevant or low-correlation features, and enhance anti-interference capabilities.

[0098] The above are only preferred embodiments of the present invention and are not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A compact ground wave radar marine vessel track prediction method, characterized in that: The following steps are involved: The AIS track data in the radar observation area is obtained and converted into the distance, speed and azimuth in the radar coordinate system as characteristic parameters, which are used as input data after preprocessing; Based on the sliding window method, the continuous time series in the input data is divided into several samples, each of which contains multi-dimensional target features of n historical time steps, which are used as the input of the prediction model; After the prediction model is trained, it outputs the prediction results of a single time step in the future and restores them to the distance, speed and azimuth corresponding to the actual physical quantities; Among them, the prediction model has N layers of stacked encoders, each encoder layer extracts time series features of different depths from the input features; each encoder layer uses several different attention mechanisms to model the time series features; each encoder layer contains a feedforward neural network, which is used to perform nonlinear transformation on the features and further extract deep time series information; the output of the last layer of encoder is converted into the prediction features of the future time steps of the output through a linear mapping layer.

2. A compact ground wave radar marine vessel track prediction method as claimed in claim 1, characterized in that: The track data obtained by the ship automatic identification system is used as the data source for model training and testing. The track data at least includes the longitude, latitude, speed and heading of the target ship.

3. A compact ground wave radar marine vessel track prediction method as claimed in claim 1, characterized in that: During the prediction model training, the longitude, latitude and speed are converted into distance, azimuth and Doppler speed in the radar coordinate system as feature parameters and normalized.

4. A compact ground wave radar marine vessel track prediction method as claimed in claim 1, characterized in that: Based on the sliding window method, the continuous time series in the input data is divided into several samples. Specifically, each track data is regarded as a continuous time series, and the sliding window method is used to slide the data sequence according to the set window length and step size. Each slide generates a new set of data, and gradually generates training and test samples.

5. A compact ground wave radar marine vessel track prediction method as claimed in claim 1, characterized in that: The prediction model is shown below: in, represents the feature vector of the predicted future time step t, Θ represents the model parameters, and X t-n:t-1 Represents the observation data matrix of n historical time steps, x t-1 represents the feature vector of the t-1th time step, and F represents the feature dimension of each time step.

6. A compact ground wave radar marine vessel track prediction method as claimed in claim 1, characterized in that: The prediction model is the DF-Transformer model. The input of the model is the preprocessed multi-time step historical track data, and the output is the single time step target prediction state.

7. A compact ground wave radar marine vessel track prediction method as claimed in claim 1, characterized in that: The data input to the prediction model is sequentially embedded, positionally encoded, multi-layered encoders, and linearly mapped to obtain output data; wherein each encoder layer includes a multi-headed attention layer and a feedforward neural network connected in sequence, and both the multi-headed attention layer and the feedforward neural network are connected to corresponding normalization layers.

8. A compact ground wave radar marine vessel track prediction method as claimed in claim 1, characterized in that: During the training of the prediction model, the global dependency between time steps is captured by calculating the attention weights of each time step in the input sequence with other time steps; each attention head models the relationship between sequence features from a different perspective, and a comprehensive representation is obtained by splicing the results of multiple heads.

9. A compact ground wave radar marine vessel track prediction method as claimed in claim 1, characterized in that: During the prediction model training, supervised training is performed using historical track samples and the corresponding target prediction values; based on the mean square error loss function, the model parameters are gradually optimized by the gradient descent method.

10. A system for implementing the compact ground wave radar marine vessel track prediction method according to any one of claims 1 to 9, characterized in that: include: The data acquisition preprocessing module is configured to: obtain the ship track data in the radar observation area, convert it into the distance, speed and azimuth in the radar coordinate system as characteristic parameters, and use the characteristic parameters as input data after preprocessing; The sliding window module is configured to: split the continuous time series in the input data into several samples based on the sliding window method, each sample contains multi-dimensional target features of n historical time steps, and use it as the input of the prediction model; The track prediction module is configured as follows: after the prediction model is trained, the prediction result of a single time step in the future is output and restored to the distance, speed and azimuth corresponding to the actual physical quantity; Among them, the prediction model has N layers of stacked encoders, each encoder layer extracts time series features of different depths from the input features; each encoder layer uses several different attention mechanisms to model the time series features; each encoder layer contains a feedforward neural network, which is used to perform nonlinear transformation on the features and further extract deep time series information; the output of the last layer of encoder is converted into the prediction features of the future time steps of the output through a linear mapping layer.

Images