Radar track classification method based on double-layer LSTM model

Abstract

The invention provides a radar track classification method based on a double-layer LSTM model. The radar track classification method is used for classifying radar targets such as unmanned ships, unmanned aerial vehicles and birds in real time. According to the method, in combination with kinematics characteristics of the variable-length radar track, a self-adaptive data preprocessing strategy is adopted, forward filling and backward truncation are carried out on the track, and the effectiveness of characteristic data is ensured through a Masking technology and standardized normalization processing. A double-layer long-short-term memory (LSTM) model is used for time sequence modeling, a mixed attention mechanism is introduced to improve the extraction capability of the model on key track information, abnormal points and non-key track points in data are effectively processed, a batch processing strategy is adopted to improve the real-time performance of the system, and the method is suitable for monitoring unmanned ships and other low-altitude targets.

Description

Technical Field

[0001] The invention belongs to the field of radar data processing. Background Art

[0002] With the widespread use of unmanned vehicles (USVs) and unmanned aerial vehicles (UAVs) in sea and air environments, traditional manual target discrimination methods are difficult to meet the real-time and accuracy requirements in complex environments. Current radar target classification methods, especially in complex low-altitude or sea environments, are often greatly limited by clutter, echo intensity differences, and the complexity of target motion trajectories. Current radar target classification technologies have the following problems:

[0003] 1. In a complex environment, clutter and environmental interference lead to increased classification errors. On the one hand, radar tracks are time series data of indefinite length. On the other hand, their target types may undergo complex dynamic switching. The same track does not necessarily always correspond to the same target type, which further increases the difficulty of classification.

[0004] 2. Traditional methods are difficult to meet the needs of real-time classification of multiple targets, especially in the detection and classification tasks of low-altitude small targets and surface targets, which require higher classification accuracy and faster processing capabilities. Summary of the invention

[0005] In order to solve the above technical problems, the present invention proposes a radar track classification method based on a double-layer LSTM model, which comprehensively considers the kinematic characteristics and historical track information of the target, and introduces the LSTM model and attention mechanism to further improve the accuracy and stability of classification.

[0006] The present invention proposes a radar track classification method based on a double-layer LSTM model. Combining the temporal characteristics of variable-length radar track data, an adaptive data preprocessing strategy is adopted to construct a double-layer LSTM neural network and utilize the global and local hybrid attention mechanism to extract the key features in the track, so as to dynamically classify low-altitude targets such as unmanned boats, drones and birds.

[0007] Furthermore, in data preprocessing, the indefinite-length track data are forward-filled and backward-truncated, filling waypoints are added and marked in the forward filling process, the filling waypoints are skipped in the standardization process, and masks are used to avoid filling interference during the training and prediction process.

[0008] Furthermore, the construction of the two-layer LSTM model includes: the first layer is used to capture short-term temporal dependencies, the second layer LSTM is configured with a smaller number of hidden units to extract the global movement and change trend of the target; the weight of a specific time step is dynamically adjusted through a global and local mixed attention mechanism, and a time distribution layer is used to process each time step in the track. The model models the input of each time step and outputs a classification result.

[0009] Furthermore, the hybrid attention mechanism includes global and local attention mechanisms, which assign different weights to the time steps of the entire sequence and specific local areas respectively. The global attention mechanism assigns a global weight to each time step t. The calculation formula of the global weight is:

[0010]

[0011] in, is a global scoring function that measures the importance of time step t in the entire sequence; the local attention mechanism assigns local weights to the local features of the radar track. The calculation formula of the local weight is:

[0012]

[0013] Where g1 to g2 are the boundaries of the local time window, represents the local score; the final mixed attention weight is the weighted sum of global weight and local weight:

[0014]

[0015] λ is the balance coefficient, which controls the relative influence of global and local attention:

[0016]

[0017] Furthermore, a batch processing mechanism is introduced in data filling, normalization, model training and prediction to process multiple target track data in parallel. The number of tracks processed each time is determined by the batch upper limit or time threshold, and a soft classification strategy is used to perform real-time classification and result output of LSTM.

[0018] The present invention is integrated into a radar target classification system, including a radar signal receiving module, a data preprocessing module, an LSTM classification module and a result output module. The system receives radar track data through a network interface and can effectively process track data of indefinite length. In the process of dynamic changes in waypoint types, the target classification results given by the system also change dynamically with the waypoints, and through batch processing strategies, the classification processing capability and real-time performance of the system are effectively improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 The model structure of LSTM and hybrid attention model.

[0020] Figure 2 Iterate training results for the model. DETAILED DESCRIPTION

[0021] The technical solution of the present invention is further explained below in conjunction with the accompanying drawings and embodiments.

[0022] The present invention proposes a radar track classification method based on a double-layer LSTM model:

[0023] The preferred data preprocessing process includes:

[0024] The indefinite-length track data are forward-filled and backward-truncated. In the forward-filling process, filling waypoints are added and marked. In the normalization process, filling waypoints are skipped. Masking is used to avoid filling interference during the training and prediction process of the neural network.

[0025] For tracks whose length is less than L, a forward filling method is used to fill the missing position with a specific value M:

[0026]

[0027] For track data whose length exceeds L, it is truncated to L time steps.

[0028] The standardization tool StandardScaler is used for normalization to eliminate dimensional differences and improve the convergence speed of the model:

[0029]

[0030] in, and They are the features x j By normalizing, the feature data can be scaled to the same scale, improving the training efficiency and prediction performance of the model.

[0031] In the preferred two-layer LSTM model: the first layer is used to capture short-term temporal dependencies, the second layer LSTM is configured with a smaller number of hidden units, extracts the global motion characteristics and change trends of the target, and dynamically adjusts the weights of specific time steps through a global and local mixed attention mechanism. The time distribution layer is used to process each time step in the track. The model models the input of each time step and outputs the classification result.

[0032] The first layer of LSTM is responsible for processing the input track data X = {x1, x2, ..., x T}, where x t Represents the input features at time step t. The hidden state of the first layer LSTM represents the output of the first LSTM layer at time step t:

[0033]

[0034] The second layer LSTM receives the output from the first layer LSTM Further model higher-level temporal features and generate new hidden states

[0035]

[0036] The preferred hybrid attention mechanism includes global and local attention mechanisms, which assign different weights to the time steps of the entire sequence and specific local regions respectively. The hybrid attention mechanism is used as a layer of the LSTM model to dynamically weight the hidden state of each time step after each time step t. By calculating the global scoring function The model assigns a global weight to each time step

[0037]

[0038] in, A measure of the importance of time step t in the entire sequence.

[0039] The local attention mechanism assigns local weights to significant changes in the radar track, such as the turning or acceleration of the drone. Local scoring functions consider only a specific local time window:

[0040]

[0041] Where g1 to g2 are the boundaries of the local time window, Indicates local rating.

[0042] The global and local attention weights are finally weighted and summed by a balance coefficient λ:

[0043]

[0044] Among them, λ controls the relative influence of global and local attention. The final classification result It is the output based on mixed attention weighting:

[0045]

[0046] where y t represents the classification output at time step t, and is the final classification result after weighting.

[0047] Preferably, a batch processing mechanism is introduced in data filling, normalization, model training and prediction, and multiple target track data are processed in parallel. The upper limit of the track processed each time is determined by the batch number or time threshold, and a soft classification strategy is used for real-time classification and result output of LSTM.

[0048] The number of tracks processed in each batch is set to B, that is, B tracks are input into the model together each time. Each track contains T time steps, and each time step contains F features. In order to handle tracks of varying lengths, all tracks are padded to a uniform length of T. max (i.e., the maximum number of time steps). When the number of targets is insufficient, it will take too long to stack the tracks to the upper limit. To ensure the real-time classification, it is necessary to limit the stacking time limit so that batch tracks can be input into the model in time.

[0049] (1) Definition of input data:

[0050] Define the input track data as X (b) , where b∈{1, 2, ..., B} represents the bth track. The data matrix of the bth track can be expressed as:

[0051]

[0052] in, is the feature vector of the tth time step, containing F features, T (b) is the number of time steps of the bth track.

[0053] (2) Filling and normalization of batch tracks:

[0054] To ensure that the time step length of each track in a batch is consistent, the track data needs to be padded to length T max The filling value is a special marker value M, and these filling values ​​are ignored by the Masking layer. The filled data is represented as:

[0055]

[0056] The normalization operation is applied to the features of each time step t Standardization is defined as:

[0057]

[0058] Among them, μ x and σ x are the mean and standard deviation of the features respectively.

[0059] (3) Batch model prediction:

[0060] All padded and normalized track data are combined into a three-dimensional tensor Represents the feature matrix of B tracks, each of which has a length of T max , each time step has F features. After batch input into the LSTM model, the prediction result of all track data is an output matrix Where C is the number of classification categories. Each element of the output It represents the target classification prediction result of the b-th track at time step t. Finally, the target classification probability of each time step is output through the Softmax activation function:

[0061]

[0062] where y t is the classification result at time step t, W y and b y are weights and biases. The loss function uses cross entropy loss:

[0063]

[0064] Among them, y tc and Represent the true label and predicted probability respectively. By minimizing the cross entropy loss, the classification effect of the model is optimized.

[0065] Example:

[0066] 1. Data collection and preprocessing:

[0067] In the implementation process of the present invention, the target track data is first collected, covering a variety of typical target types (including unmanned boats, drones, birds, ships, etc.). Each line of each track file represents a waypoint, and the target type is marked at each waypoint to cope with the target type switching of different waypoints in the same track.

[0068] The collected features include: echo amplitude σ, timestamp t, distance d, azimuth θ, elevation angle φ, heading γ, speed v, altitude h, and the rate of change of each kinematic feature. The calculation formula is:

[0069]

[0070] These rates of change are calculated by taking the differences between the adjacent waypoints and the time difference Δt, where the first waypoint does not contain a rate of change.

[0071] 2. Model construction and training:

[0072] The present invention adopts a double-layer LSTM structure to process the temporal features in the track, and combines the hybrid attention mechanism, the Dropout layer and the TimeDistributed layer to improve the robustness of the model and the ability to process the output of each time step.

[0073] The first LSTM layer contains 128 hidden units, which are responsible for extracting shallow temporal features, followed by a Dropout (0.5) layer to reduce overfitting. The hidden state h output by each LSTM layer is t is passed to the next layer.

[0074] The second LSTM layer contains 64 hidden units to further process deep temporal features, and is also followed by a Dropout (0.2) layer to further prevent overfitting.

[0075] The hybrid attention mechanism is embedded after the two-layer LSTM and the weight of each time step is calculated through global and local attention.

[0076] Mixed Attention Weights is the weighted sum of global and local weights:

[0077]

[0078] The output of the hybrid attention mechanism is weighted by the classification output y at each time step t t .

[0079] The TimeDistributed layer is used to process the hidden state of each time step step by step, pass the output of each time step to a fully connected layer, and calculate the classification result of each time step through the Softmax activation function. To ensure the accuracy of classification, the target type of the first waypoint is not judged. At least start judging from the second waypoint.

[0080] 3. Model optimization and training:

[0081] Use the Adam optimizer and set the learning rate to 0.0005:

[0082] optimizer=Adam(lr=0.0005) (16)

[0083] The loss function is cross entropy loss, which is used to optimize the classification effect of the model.

[0084] During model training, the training rounds are 150 and 10% of the validation set is used for model validation.

[0085] 4. Performance Evaluation

[0086] The model was trained and tested by selecting track data of various target categories including unmanned boats, drones, birds, etc. to verify its effectiveness. In order to compare the differences between the LSTM model and the traditional model in classification accuracy and processing efficiency, the model based on CNN (convolutional neural network) was also evaluated. The following is a detailed description of the experimental results.

[0087] (1) Comparison of classification accuracy: During the iterative training of the model, the CNN-based model has obvious limitations in classification accuracy. Specifically, the loss value of the CNN model is difficult to converge during the training process. The final accuracy on the training set is 95%, but the accuracy on the test set is only about 80%, showing an obvious overfitting problem. The LSTM model achieves an accuracy of 95% and 92% on the training set and test set, respectively, and the loss value converges well, indicating that it has significant advantages in processing time series data.

[0088] (2) Comparison of model processing efficiency: In terms of processing efficiency, the advantage of CNN lies in its fast single prediction speed. On a processing platform with an i7-1260p CPU and 16G memory, the average prediction time is about 12 milliseconds, while the average prediction time of LSTM is 50 milliseconds, and about 20 radar track data can be processed per second. Although the single prediction time of LSTM is relatively long, by introducing a batch processing strategy, for example, for a radar with a 2-second data rate, the batch upper limit is B, and the processing capacity per cycle can be increased to 40×B, which greatly improves the parallel processing capability of the system.

[0089] (3) Classification performance of unmanned boats: In order to verify the effectiveness of the model in classifying unmanned boat targets, actual collected ship track data and simulated unmanned boat data were used. The echo amplitude of unmanned boats on the sea surface is significantly higher than that of birds, and its height change is relatively stable. The experimental results show that the LSTM model can effectively distinguish between unmanned boats and bird targets, and its classification accuracy reaches more than 92%. After introducing the hybrid attention mechanism, the model is more sensitive to the classification of unmanned boats in significant actions such as turning or accelerating, further improving the classification accuracy.

[0090] (4) Dynamic classification performance: The following table shows the dynamic changes in the bird track data. The first waypoint does not make a classification decision, and the target type given is null. From the second waypoint to the fifth waypoint, the target flies stably and is classified as a bird. However, the echo amplitude of the sixth waypoint is 0, indicating that the radar does not detect the point track. This waypoint is an extrapolated waypoint, and the target speed increases significantly. The LSTM model no longer classifies the target as a bird. The extrapolated distance and speed of the seventh waypoint are abnormal. The waypoint has been associated with other clutter and is no longer valid. The model gives the classification result "Else".

[0091]

[0092] The radar track classification method based on the LSTM model of the present invention has high accuracy and good real-time performance when dealing with the classification tasks of unmanned boats, drones, birds and other targets. Compared with the traditional convolutional neural network model, this method has advantages in dealing with time series data and dynamic changes of targets.

Claims

1. A radar track classification method based on a double-layer LSTM model, characterized by: Combined with the temporal characteristics of variable-length radar track data, an adaptive data preprocessing strategy was adopted. By constructing a two-layer LSTM neural network, the global and local hybrid attention mechanism was used to extract the key features in the track, and the unmanned boats, drones, and birds were dynamically classified.

2. The radar track classification method based on the double-layer LSTM model according to claim 1, characterized in that: In data preprocessing, the indefinite-length track data are forward-filled and backward-truncated. In the forward filling process, filling waypoints are added and marked. In the standardization process, filling waypoints are skipped. Masking is used to avoid filling interference during training and prediction.

3. The radar track classification method based on the double-layer LSTM model according to claim 1, characterized in that: The construction of the two-layer LSTM model includes: the first layer is used to capture short-term temporal dependencies, the second layer LSTM is configured with a smaller number of hidden units to extract the global movement and change trend of the target; the weight of a specific time step is dynamically adjusted through a global and local mixed attention mechanism, and a time distribution layer is used to process each time step in the track. The model models the input of each time step and outputs a classification result.

4. The radar track classification method based on the double-layer LSTM model according to claim 1, characterized in that: The hybrid attention mechanism includes global and local attention mechanisms, which assign different weights to the time steps of the entire sequence and specific local areas respectively. The global attention mechanism assigns a global weight to each time step t. The calculation formula of the global weight is: in, is a global scoring function that measures the importance of time step t in the entire sequence; the local attention mechanism assigns local weights to the local features of the radar track. The calculation formula of the local weight is: Where g1 to g2 are the boundaries of the local time window, represents the local score; the final mixed attention weight is the weighted sum of global weight and local weight: λ is the balance coefficient, which controls the relative influence of global and local attention:

5. The radar track classification method based on the double-layer LSTM model according to claim 1, characterized in that: A batch processing mechanism is introduced in data filling, normalization, model training and prediction to process multiple target track data in parallel. The number of tracks processed each time is determined by the batch upper limit or time threshold. A soft classification strategy is used for real-time classification and result output of LSTM.

Images