Radar target identification method

Abstract

The invention relates to a radar target identification method. The method comprises the following steps: acquiring an RCS sequence and counting input RCS samples; performing feature extraction on the RCS sequence by using a first feature extraction model and a second feature extraction model to obtain a first feature and a second feature; by taking the time step length of the second feature extraction model as a window, performing feature alignment on the first feature and the second feature to obtain an aligned first feature as a third feature; based on an attention mechanism, performing feature fusion on the second feature and the third feature to obtain a fused feature; and when the RCS sample count does not exceed a first set value, predicting a radar target recognition result according to the first feature, otherwise, predicting the radar target recognition result according to the fusion feature. According to the invention, the radar target can be accurately identified, and the method has high universality.

Description

Technical Field

[0001] This invention relates to the field of radar target recognition technology, and in particular to a radar target recognition method. Background Technology

[0002] Radar target identification plays a vital role in both military and civilian fields and has broad application prospects. Based on the degree of target identification, radar target identification can be divided into the following three levels: detection, used to separate the target from the background; classification, used to determine the type of target, such as whether the target is a rocket, missile, or space station; and identification, used to determine the type of target, such as whether a rocket is the main body or debris, or a missile is a warhead or fragments.

[0003] In different scenarios, there are various algorithms available for radar target recognition, each with its own shortcomings:

[0004] Logistic regression: cannot handle non-linear relationships, relies heavily on feature engineering, requires high independence between features, and is sensitive to outliers;

[0005] K-Nearest Neighbors (KNN) algorithm: computationally intensive, performs poorly on high-dimensional data (curse of dimensionality), requires determining a suitable value for K, and is sensitive to noise;

[0006] Support Vector Machine (SVM): It is difficult to implement with large-scale training samples, has difficulty solving multi-class classification problems, and is sensitive to missing data and parameters;

[0007] Decision trees are prone to overfitting, are not accurate enough for certain types of data, and are too sensitive to small changes, resulting in the generation of completely different trees.

[0008] Random forests may overfit on some noisy classification or regression problems. They may not be as effective as other simpler models for small datasets or low-dimensional data, and training and prediction are relatively slow.

[0009] Therefore, a radar target identification method is needed that can accurately identify radar targets and has high versatility. Summary of the Invention

[0010] The technical problem to be solved by the present invention is to provide a radar target identification method that can accurately identify radar targets and has high versatility.

[0011] The technical solution adopted by this invention to solve its technical problem is: to provide a radar target identification method, comprising the following steps:

[0012] Obtain the RCS sequence and count the input RCS samples;

[0013] The RCS sequence is subjected to feature extraction using a first feature extraction model and a second feature extraction model, respectively, to obtain the first feature and the second feature.

[0014] Using the time step of the second feature extraction model as a window, the first feature and the second feature are aligned, and the aligned first feature is used as the third feature.

[0015] Based on the attention mechanism, the second and third features are fused to obtain the fused features;

[0016] When the RCS sample count does not exceed the first set value, the radar target identification result is predicted based on the first feature; otherwise, the radar target identification result is predicted based on the fused features.

[0017] Furthermore, the step of aligning the first feature with the second feature extraction model using the output time of the second feature extraction model as the time window includes:

[0018] Using the time step of the second feature extraction model as a window, all first features extracted by the first feature extraction model within the current window are obtained and aggregated.

[0019] Furthermore, the step of acquiring and aggregating all first features extracted by the first feature extraction model within the current window includes:

[0020] Retrieve all first features extracted by the first feature extraction model within the current window;

[0021] The weights of the first feature at each time step of the first feature extraction model are learned using an attention mechanism.

[0022] Aggregate all the first features based on the obtained weights.

[0023] Furthermore, the feature fusion based on the attention mechanism for the second and third features includes:

[0024] The weights of the second and third features are learned using an attention mechanism;

[0025] The second and third features are fused based on the obtained weights.

[0026] Furthermore, the learning of the weights of the second and third features using the attention mechanism is achieved by treating the second feature as K and V in the multi-head attention mechanism and the third feature as Q in the multi-head attention mechanism.

[0027] Furthermore, the radar target recognition result is obtained by predicting based on the input first feature or fused feature using a linear classification layer.

[0028] Furthermore, the first feature extraction model is a TCN network.

[0029] Furthermore, the second feature includes at least one of the following: processing location feature parameters, scattering feature parameters, distribution feature parameters, and transformation feature parameters.

[0030] Beneficial effects

[0031] Due to the adoption of the above technical solutions, the present invention has the following advantages and positive effects compared with the prior art: The present invention uses a first feature extraction model based on motion features to extract the first feature, and a second feature extraction model based on RCS sequence features to extract the second feature. Then, the two features are aligned and fused using an attention mechanism, and the radar target recognition result is predicted based on the fused feature, which improves the recognition accuracy, precision and versatility. In addition, the dynamic selection mechanism is adopted to select whether to make prediction based on the first feature or the fused feature, which can take into account both the real-time performance and high accuracy of the recognition model. Attached Figure Description

[0032] Figure 1 This is a flowchart of an embodiment of the present invention;

[0033] Figure 2 This is a schematic diagram of the causal convolution structure according to an embodiment of the present invention;

[0034] Figure 3 This is a schematic diagram of the dilated convolution structure according to an embodiment of the present invention;

[0035] Figure 4 This is a schematic diagram of the residual block structure of the TCN network according to an embodiment of the present invention;

[0036] Figure 5 This is a schematic diagram of the attention mechanism according to an embodiment of the present invention. Detailed Implementation

[0037] The present invention will be further illustrated below with reference to 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. Furthermore, it should be understood that after reading the teachings of this invention, those skilled in the art can make various alterations or modifications to the invention, and these equivalent forms also fall within the scope defined by the appended claims.

[0038] The embodiments of the present invention relate to a radar target identification method, such as... Figure 1 As shown, it includes the following steps:

[0039] Obtain the RCS sequence and count the input RCS samples;

[0040] The RCS sequence is subjected to feature extraction using a first feature extraction model and a second feature extraction model, respectively, to obtain the first feature and the second feature.

[0041] Using the time step of the second feature extraction model as a window, the first feature and the second feature are aligned, and the aligned first feature is used as the third feature.

[0042] Based on the attention mechanism, the second and third features are fused to obtain the fused features;

[0043] When the RCS sample count does not exceed the first set value, the radar target identification result is predicted based on the first feature; otherwise, the radar target identification result is predicted based on the fused features.

[0044] The first feature extraction model can be a Temporal Convolutional Network (TCN) or a Recurrent Neural Network (RNN), such as Long Short Term Memory (LSTM) and Gate Recurrent Unit (GRU), or any algorithm for temporal modeling.

[0045] This implementation employs a temporal convolutional network (TCN), which utilizes convolutional methods and outperforms other models on certain tasks. While recurrent architectures theoretically possess the ability to capture infinitely long memories, TCN exhibits an even longer memory, making it more suitable for domains requiring long-term historical data. The TCN architecture is not only more accurate, simpler, and clearer than canonical recurrent networks such as LSTM and GRU, but more importantly, the convolutions in its architecture are causal compared to other temporal models, meaning there is no information "leakage" from the future to the past. Furthermore, the method emphasizes the use of a combination of very deep networks (enhanced with residual layers) and dilated convolutions to construct very long effective historical scales (i.e., the network's ability to make predictions about the past).

[0046] TCN is based on two principles: the output produced by the network must be of the same length as the input, and there must be no leakage from the future to the past. To achieve the first point, TCN uses a one-dimensional fully convolutional network (FCN) architecture, where each hidden layer is the same length as the input layer, with zero padding added to ensure that subsequent layers are the same length as previous layers. To achieve the second point, TCN uses causal convolution, meaning that the output at time t is only convolved with elements at time t and earlier elements from the previous layer, such as... Figure 2 As shown. Therefore, simply put: TCN = 1D FCN + causal convolutions.

[0047] A major drawback of this basic design is that achieving long-term effective history scale requires a very deep network or very large filters, neither of which is feasible. Therefore, dilated convolutions emerged. A simple causal convolution can only rewind a history of linear size within the network depth. This makes applying the aforementioned causal convolutions challenging for sequence tasks, especially those requiring longer histories. Dilation is equivalent to introducing a fixed stride between every two adjacent filter taps. When d=1, the dilated convolution reduces to a regular convolution. Using a larger expansion allows the output to represent a wider range of inputs, effectively expanding the receptive field of the convolutional neural network. Dilated convolutions, such as... Figure 3 As shown.

[0048] However, even with dilated causal convolutions, the model can sometimes still be very deep. Deep network structures can cause problems like vanishing gradients. To address this, a structure similar to residual blocks in ResNet is used, making the TCN structure more generalizable. In standard ResNet, the input is directly added to the output of the residual function, while in TCN, the input and output may have different widths. To account for this difference in input and output widths, an additional 1x1 convolution is used to ensure that element-wise addition receives tensors of the same shape. The residual structure is as follows: Figure 4 As shown.

[0049] The TCN model has the following advantages:

[0050] (1) Parallelism: In RNNs, predictions for later time steps must wait for their preceding predictions to complete. Convolutions can be performed in parallel because each layer uses the same filters. Therefore, during training and evaluation, a long input sequence can be processed as a whole in a TCN, rather than sequentially as in an RNN.

[0051] (2) Flexible receptive field: A TCN can change the size of its receptive field in a variety of ways. For example, stacking more dilated (causal) convolutional layers, using a larger dilation factor, or increasing the size of the filters (kernel size) are all viable options. Therefore, TCNs can better control the memory size of the model and easily adapt to different domains.

[0052] (3) Stable gradients: Unlike recurrent neural networks (RNNs), TCNs have a backpropagation path that is different from the temporal direction of the sequence. Therefore, TCNs can avoid the gradient explosion / vanishing problem, which is a major problem of RNNs (which led to the emergence of LSTM, GRU, HF-RNN, etc.).

[0053] (4) Low memory requirements for training: Especially with long input sequences, LSTM and GRU can easily consume a lot of memory to store partial results from multiple unit gates. However, in TCN, filters are shared across layers, and the backpropagation path depends only on the network depth. Therefore, in practice, we have found that gated RNNs may use more memory than TCNs.

[0054] (5) Variable-length input: Just as RNNs model variable-length inputs in a cyclic manner, TCNs can also obtain inputs of arbitrary lengths by sliding a one-dimensional convolutional kernel. This means that TCNs can serve as an alternative to RNNs for arbitrary-length sequence data.

[0055] The model has the following drawbacks:

[0056] (1) During evaluation / testing, RNNs only need to maintain a hidden state and accept a current input to generate a prediction. In other words, the "summary" of the entire history is provided by a fixed-length set of vectors, which can discard the actually observed sequence. In contrast, TCNs need to receive a valid history length from the original sequence, and therefore may require more memory during evaluation;

[0057] (2) Potential parameter changes during domain migration: Different domains may have different requirements for the amount of history that the model needs to predict. Therefore, when the model is moved from a domain that requires less memory (i.e., smaller k and d) to a domain that requires more memory (i.e., larger k and d), TCN may perform poorly due to the lack of a sufficiently large receptive field.

[0058] To address the aforementioned issues of the TCN model, a second feature extraction model based on RCS sequence features is introduced for multimodal fusion recognition. This model is implemented using an RCS network. The RCS network extracts RCS features (i.e., the second feature) from the input RCS sequence. In this embodiment, the RCS features include one or more of the following parameters: processing location feature parameters, scattering feature parameters, distribution feature parameters, and transformation feature parameters.

[0059] Location feature parameters describe the average and specific locations of the target RCS time series, including mean, median, maximum, and minimum values;

[0060] The dispersion characteristic parameters describe the dispersion of the target RCS time series across the entire real axis, including range, variance, standard deviation, and coefficient of variation.

[0061] Distribution characteristic parameters describe the graphical characteristics of the overall density function of the target RCS statistical distribution, including skewness coefficient and kurtosis coefficient;

[0062] The sequence is transformed by changing the feature parameters, performing Fourier transform, wavelet transform, etc., and then performing feature extraction processing on the transformed sequence.

[0063] Let {x} i The sequence of target RCS measured by radar within a selected time period is represented by the group i = 1, 2, ..., N. The method for calculating the characteristics of each RCS data within this time period is as follows:

[0064] RCS mean:

[0065]

[0066] RCS median: for the RCS sequence {x i Sort the numbers i = 1, 2, ..., N by size, and the middle number is the median.

[0067] RCS range: the difference between the maximum and minimum values.

[0068] RCS variance:

[0069]

[0070] The RCS (Relative Standard Deviation) coefficient of variation is the ratio of the absolute value of the standard deviation to the absolute value of the arithmetic mean.

[0071]

[0072] The RCS skewness coefficient is the ratio of the third-order central moment to the cube of the standard deviation:

[0073]

[0074] The RCS kurtosis coefficient is the ratio of the fourth-order central moment to the fourth power of the standard deviation:

[0075]

[0076] Extracting the entropy of the RCS: Extracting the entropy of the RCS sequence {x} i ,x i ≤x i+1 Divide the {i = 1, 2, ..., N-1} into K intervals:

[0077] {[K i ,K i+1 ], i = 1, 2, ..., K}

[0078] Calculate the number p of RCS sequence values ​​falling into each interval. i For i = 1, 2, ..., K, the formula for calculating the entropy information of RCS is as follows:

[0079]

[0080] Extracting RCS periodic information. The target RCS sequence is processed using a Fourier transform, and the first half of the Fourier-transformed RCS sequence is used to obtain the frequency domain representation of the RCS {z}. i ,i=1,2,…,[N / 2]}, extract the position of the maximum value, the position of the second maximum value and the position of the third maximum value of the sequence. This information reflects the micro-motion characteristics of the target such as spin, precession and nutation.

[0081] Fourier transform spectrum. Perform a Fourier transform on the RCS sequence, and let the resulting truncated sequence be {y}. i The absolute value and average value of the sequence are used as estimates of the cosine transform eigenvalues ​​of the RCS sequence.

[0082] Low-frequency energy ratio. Perform a Fourier transform on the RCS sequence, and let the resulting truncated sequence be {y}. i ,i=1,2,…,M}, select {y i The sequence i = 1, 2, ..., K = M / / 4 is a low-frequency sequence. The formula for calculating the low-frequency energy ratio is as follows:

[0083]

[0084] Calculating a feature vector from the entire RCS sequence involves too little data and is unstable. In some preferred implementations, a hierarchical sequential input strategy can be used for identification. For example, assuming the RCS sequence length is 100, features are first calculated using a sliding window with a window length of 10 and a sliding step size of 1. The dimension of the feature vector calculated each time is 11, and the features include: mean, maximum, minimum, median, 90th percentile, range, variance, standard deviation, coefficient of variation, skewness coefficient, and kurtosis coefficient. 100 measurement points are used as input, and a window length of 10 is selected in the sequential sliding window processing.

[0085] The core of multimodal fusion recognition is to utilize the outputs of the TCN network and RCS network, and perform feature fusion through an attention mechanism to ultimately achieve target classification. Specifically, it includes the following steps:

[0086] Using the time step of the RCS network as a window, the first feature and the second feature are aligned, and the aligned first feature is used as the third feature.

[0087] Based on the attention mechanism, the second and third features are fused to obtain the fused features.

[0088] To illustrate the feature fusion method, we take a real-world deployment task as an example. We set the output frequency of the TCN network to be higher (every 5 measurements) and the output frequency of the RCS network to be lower (every 100 measurements).

[0089] First, the time window needs to be aligned with the output time of the RCS network: within every 100 measurements, the TCN network outputs 20 recognition results. These 20 results from the TCN network within the 100-measurement window are then aggregated (using an attention mechanism) to obtain a feature representation corresponding to the RCS network result.

[0090] Then, an attention mechanism is used to learn the importance of features at different time steps, and the 20 motion feature TCN feature vectors within 100 measurement windows are aggregated into a single feature vector F. velocity The RCS network outputs a feature vector F after 100 measurements. RCS The attention mechanism is used to learn the weights of features from different modalities. Specifically, a linear projection network is used to transfer the weights of F... velocity and F RCS Projected onto the same dimension, F velocity As Q and F in the attention mechanism RCS As K and V in the attention mechanism, attention scores are learned through a multi-head attention mechanism to learn the final fused feature representation and capture complementary information between the two modalities.

[0091]

[0092] Finally, the fused attention feature vectors are passed through a linear classification layer to obtain classification results for several target classes.

[0093] More specifically, the attention mechanism used in this embodiment is as follows: Figure 5 As shown. The formula for calculating attention weights is...

[0094]

[0095] Where Q is the query matrix, K is the key matrix, and V is the value matrix. It represents the dimension of the key, used to scale the dot product result to avoid it becoming too large.

[0096] Multi-head attention involves dividing Q, K, and V into multiple h heads, computing the attention in parallel, and then concatenating and linearly transforming the results.

[0097] MultiHead(Q,K,V)=Concat(head1,head2,...,head h W o

[0098] head i =Attention(Q) i ,K i V i ), It is a linear transformation matrix that maps the splicing result back to the output space.

[0099] The final output is further processed through addition and layer normalization:

[0100] Z=LayerNorm(Attention(Q,K,V)+X)

[0101] LayerNorm is a layer normalization method used to stabilize training.

[0102] By fine-tuning on a training set of a certain model, the fusion network achieved an accuracy of 96.429% in recognizing 7 types of targets and 99.702% in recognizing 6 types of targets for a certain task, showing a significant improvement in recognition accuracy.

[0103] Meanwhile, to balance the real-time performance and high accuracy of the recognition model, a dynamic selection mechanism is adopted. In the real-time target recognition stage, a motion feature-based TCN network is first used for recognition. This network only requires data from 5 measurement points of the target to identify it, and it has high accuracy. When a batch of target data accumulates to 100 measurement points, the fusion recognition network can be activated. This network adaptively weights the target's motion features and RCS sequence features through an attention mechanism, which has higher recognition accuracy and is suitable for target recognition after target separation.

Claims

1. A radar target identification method, characterized in that, Includes the following steps: Obtain the RCS sequence and count the input RCS samples; The RCS sequence is subjected to feature extraction using a first feature extraction model and a second feature extraction model, respectively, to obtain the first feature and the second feature. Using the time step of the second feature extraction model as a window, the first feature and the second feature are aligned, and the aligned first feature is used as the third feature. Based on the attention mechanism, the second and third features are fused to obtain the fused features; When the RCS sample count does not exceed a first set value, the radar target classification result is predicted based on the first feature; otherwise, the radar target classification result is predicted based on the fused features.

2. The method according to claim 1, characterized in that, The step of aligning the first feature with the second feature extraction model using the output time of the second feature extraction model as the time window includes: Using the time step of the second feature extraction model as a window, all first features extracted by the first feature extraction model within the current window are obtained and aggregated.

3. The method according to claim 2, characterized in that, The step of obtaining and aggregating all first features extracted by the first feature extraction model within the current window includes: Retrieve all first features extracted by the first feature extraction model within the current window; The weights of the first feature at each time step of the first feature extraction model are learned using an attention mechanism. Aggregate all the first features based on the obtained weights.

4. The method according to claim 1, characterized in that, The feature fusion based on the attention mechanism for the second and third features includes: The weights of the second and third features are learned using an attention mechanism; The second and third features are fused based on the obtained weights.

5. The method according to claim 4, characterized in that, The method of learning the weights of the second and third features using the attention mechanism is achieved by treating the second feature as K and V in the multi-head attention mechanism and the third feature as Q in the multi-head attention mechanism.

6. The method according to claim 1, characterized in that, The radar target recognition result is obtained by predicting based on the first or fused features of the input using a linear classification layer.

7. The method according to claim 1, characterized in that, The first feature extraction model is a TCN network.

8. The method according to claim 1, characterized in that, The second feature includes at least one of the following: processing location feature parameters, scattering feature parameters, distribution feature parameters, and transformation feature parameters.

Images