A radar echo recognition method for low-altitude, slow-moving and small targets based on small samples

Abstract

The present invention discloses a method for identifying radar echoes of low, slow and small targets based on small samples, comprising the following steps: constructing a radar echo identification network model, where the radar echo identification network model includes a residual network feature extraction module, a memory enhancement module integrating a multi-head attention mechanism of a bidirectional long short-term memory network, and a similarity measurement module; training, validating and testing the radar echo identification network model using a data set.

Description

Technical Field

[0001] The present invention relates to radar echo signal processing technology

[0002] The present invention relates to a radar echo recognition method for low, slow and small targets based on small samples. Background Art

[0003] Radar detects, identifies and tracks targets by emitting electromagnetic waves, and is an important means of obtaining target information. Radar echo target recognition technology, as a technology that determines the properties and characteristics of targets through scattered echo analysis, has a very wide range of applications. The "low, slow and small" targets targeted by this technology mainly refer to slow drones with a flight altitude generally below 1,000 meters, or small unmanned ships on the water. These devices have a small radar reflection area and are difficult to detect, capture, handle and deal with, posing a great threat to the safety of important targets. The identification of low, slow and small targets is one of the important tasks of radar in the new era.

[0004] With the continuous development of technology and the continuous improvement of computing power, artificial intelligence technology has also made great progress in various fields, especially in images, text, voice, signals and many other aspects, making significant contributions to global economic development. Deep learning technology mostly relies on a large amount of labeled data to support it, but many scenarios cannot provide sufficient samples for learning due to the uniqueness of the scenario, and in practice it is not possible to collect enough data for learning. Summary of the invention

[0005] The technical problem to be solved by the present invention is that new low, slow and small targets such as unmanned boats, floating mines, submarine telescopes, and rotor drones have the problems of small target scattering area, low flight altitude, slow speed, unclear Doppler frequency shift, and low target signal-to-noise ratio during radar detection. The present invention provides a small sample low, slow and small target radar echo recognition method to solve the above problems. The method can be applied to learning scenarios where traditional neural networks cannot effectively extract low, slow and small target features under environmental constraints, and can effectively improve the radar's recognition ability of low, slow and small targets in complex environments.

[0006] The present invention is achieved through the following technical solutions:

[0007] A method for identifying low, slow and small target radar echoes based on small samples includes the following steps:

[0008] A radar echo recognition network model is constructed, which includes a residual network feature extraction module, a memory enhancement module integrating a multi-head attention mechanism of a bidirectional long short-term memory network, and a similarity measurement module; and a data set is used to train, verify and test the radar echo recognition network model.

[0009] Further preferably, the residual network feature extraction module is constructed by stacking 3 residual structures to form a 10-layer residual network for feature extraction.

[0010] Further preferably, the memory enhancement module of the fusion bidirectional long short-term memory network with multi-head attention mechanism organically combines the multi-step bidirectional long short-term memory network (bi-LSTM) and the multi-head attention mechanism. The common features are extracted through the multi-step bidirectional long short-term memory network, and the output of the multi-step bidirectional long short-term memory network is used as the input of the multi-head attention mechanism. The multi-head attention mechanism extracts the common semantic information between samples by focusing on the common features between samples.

[0011] Further preferably, the multi-step bidirectional long short-term memory function is shown as follows:

[0012]

[0013] where f and g represent neural networks, S represents the support set, K represents the number of iteration steps, represents the target feature.

[0014] Further preferably, the multi-head attention mechanism formula is shown as follows:

[0015] MultiHead(Q,K,V)=Concat(z 1 ,...,z h )w z ;

[0016] z i =Attention(QW i Q ,KW i K ,VW i V );

[0017] where, d k =d v =d model / h, where h is the number of heads of self-attention.

[0018] Further preferably, the similarity measurement module is used to learn the distance or difference between data, and describe the similarity between samples; extract features through a convolutional neural network and map the samples to a high-dimensional metric space, and measure the similarity between samples in the high-dimensional metric space; the measurement means include Manhattan distance, Euclidean distance, Mahalanobis distance, Chebyshev distance and cosine distance.

[0019] Further preferably, the goal of the convolutional neural network is to overcome the imbalance between easy and difficult samples, maximize the probability of predicting the test data set, and minimize the loss function. The target loss function is as follows:

[0020] FL(p t )=-α t (1-p t ) γ log(p t )

[0021] Among them, α t is the weight parameter, p t is the predicted value, and γ is an adjustable factor. Among them, for samples with a relatively large probability (1-p t ) γ tends to 0, and its loss function value can be reduced. For difficult samples with a relatively low true probability, (1-p t ) γ has a relatively small impact on the loss function.

[0022] Further preferably, before constructing the radar echo recognition network model, the following steps are also included:

[0023] Collect radar echo data and construct a database of radar echoes of low, slow, and small targets; after preprocessing the collected radar echo data, construct a data set including a small training set, a validation set, and a test set; the radar echo data includes the type of the target and the coordinates, size, attitude angle, and speed of the target.

[0024] Further preferably, the preprocessing includes the following steps:

[0025] First, delete invalid data, duplicate data, handle missing values, and outliers in the constructed database;

[0026] Secondly, construct a source task data set and a target task data set; and intercept the data, and the intercepted signal length is L; use the absolute amplitude detection method to detect the target in the cleaned data, set the threshold to ±0.75, and then perform target interception; among them, the interception length L

[0027] Finally, extract features from the labeled samples; the feature extraction methods include wavelet transform, frequency domain feature extraction, and feature fusion.

[0028] Further preferably, by constructing a small sample labeled data set M((x i ,y i )∈M), input an unknown radar echo signal The output classification label is:

[0029]

[0030] Among them, f is the multi-head attention kernel function trained in step S4, k is the number of categories of the dataset M, and x i represents the labeled sample in the dataset M, and y i represents the label of x i .

[0031] The present invention has the following advantages and beneficial effects:

[0032] The method provided by the present invention can effectively solve the problems of difficult identification of low, slow, and small targets, such as small scattering area, low flight altitude, slow speed, insignificant Doppler frequency shift, and low target signal-to-clutter ratio. It has the characteristics of practicability and high efficiency, and is especially suitable for learning scenarios with few target samples and difficult feature extraction in complex environments.

[0033] 1. Traditional deep learning classification and recognition algorithms cannot effectively extract the essential features of low, slow, and small targets under complex environments and insufficient samples, and the recognition difficulty is large. The method provided by the present invention can extract the essential features of the target with only a small number of samples through the attention mechanism, overcoming the deficiency of the traditional deep learning method in extracting the features of low, slow, and small targets in complex environments. It has high practicability and efficiency, can save a large amount of manpower and material resources, realize low-cost and high-flexibility operation, and the obtained verification results show that the accuracy rate > 98%;

[0034] 2. For complex application scenarios, such as in the fingerprint recognition of communication radiation sources, under the actual complex electromagnetic environment conditions, it is difficult for people to extract the essential features of the radiation source for low signal-to-noise ratio electromagnetic radiation sources, and the traditional machine learning classification and recognition algorithms are limited in this case. The method provided by the present invention is especially suitable for solving engineering application problems in complex scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] The drawings described herein are used to provide a further understanding of the embodiments of the present invention, form a part of this application, and do not limit the embodiments of the present invention. In the drawings:

[0036] Figure 1 is a flowchart of a method for radar echo recognition of low, slow, and small targets based on small samples according to the present invention;

[0037] Figure 2 is a flowchart of the model construction algorithm according to the present invention;

[0038] Figure 3 is a schematic structural diagram of the residual network of the feature extraction module according to the present invention;

[0039] Figure 4 is a schematic structural diagram of the multi-head attention enhancement module according to the present invention;

[0040] Figure 5Schematic diagram of the model training process of the present invention;

[0041] Figure 6 For the experimental results of the differences in training samples. Specific implementation manners

[0042] To make the objectives, technical solutions and advantages of the present invention clearer and more understandable, the present invention will be further described in detail below in conjunction with embodiments and the accompanying drawings. The illustrative embodiments of the present invention and their descriptions are only used to explain the present invention and do not limit the present invention.

[0043] In the following description, a large number of specific details are set forth in order to provide a thorough understanding of the present invention. However, it is obvious to those of ordinary skill in the art that the present invention does not have to adopt these specific details. In other embodiments, well-known structures, circuits, materials or methods are not specifically described in order to avoid obscuring the present invention.

[0044] Throughout the specification, the reference to "one embodiment", "embodiment", "one example" or "example" means that the specific features, structures or characteristics described in connection with the embodiment or example are included in at least one embodiment of the present invention. Thus, the phrases "one embodiment", "embodiment", "one example" or "example" appearing throughout the specification do not necessarily all refer to the same embodiment or example. In addition, the specific features, structures or characteristics can be combined in any appropriate combination and / or sub-combination in one or more embodiments or examples. In addition, those of ordinary skill in the art should understand that the drawings provided herein are for illustrative purposes only and are not necessarily drawn to scale. The term "and / or" used herein includes any and all combinations of one or more of the related listed items.

[0045] Embodiment 1

[0046] This embodiment provides a method for identifying radar echoes of low, slow and small targets based on small samples, and the specific steps are as follows:

[0047] Step 1: Collect data

[0048] Collect the radar echo signals of unmanned aerial vehicles and unmanned ship equipment of different models, number them respectively, and establish a sample library. The radar echo signals include, but are not limited to, information such as the type of target, coordinates, size, attitude angle, and speed.

[0049] Step 2: Preprocess the data in the sample library to construct a training set, a validation set and a test set

[0050] Data preprocessing mainly includes steps such as data cleaning, data annotation and feature extraction.

[0051] The purpose of data cleaning is to delete invalid data and duplicate data in the original sample library, and process missing values and outliers. The cleaned data is subjected to target detection using the absolute amplitude detection method, with the threshold set to ±0.75 (after normalization), and then the target signal is intercepted, where the interception length is L. The intercepted data is labeled, and the label is the target category manually marked in combination with expert experience. Feature extraction is performed on the classified signals and saved. The extraction methods include but are not limited to short-time Fourier transform and wavelet transform. In this implementation, discrete wavelet transform is used.

[0052] Dataset division: The preprocessed dataset in this implementation contains 11 categories. Randomly select 3 categories as the training set. The training set is divided into two cases. One is 30 samples per category, a total of 90 samples; the other is 100 samples per category, a total of 300 samples. Randomly select 3 categories with 300 samples per category as the validation set, and the remaining 5 categories as the test set. The categories of the training set, validation set, and test set do not overlap.

[0053] Step 3: Construct a small sample recognition network model for radar echoes

[0054] The small sample recognition network model is composed of three major modules: a residual network feature extraction module, a memory enhancement module that integrates the multi-step long short-term memory network's multi-head attention mechanism, and a similarity measurement module. The input data is first subjected to high-dimensional feature extraction through the residual network, then processed by the multi-head attention mechanism that integrates the Bi-LSTM to enhance performance, and finally processed by the similarity measurement module to output the classification result. The specific introduction of each module is as follows:

[0055] (1) The residual network feature extraction module is constructed by stacking 3 residual structures to form a 10-layer residual network for feature extraction. The residual network feature extraction module can ensure the learning ability of target characteristics while avoiding overfitting and degradation, and can fit the complex radar target feature extraction process as much as possible, thereby improving the accuracy of target recognition. As Figure 3 shown.

[0056] (2) The memory enhancement module that integrates the multi-step long short-term memory network's multi-head attention mechanism combines the multi-step bidirectional long short-term memory network (bi-LSTM) and the multi-head attention mechanism to learn the common features between samples. On the basis of the strongly representative feature extraction module, full conditional encoding is added to enhance the common feature extraction performance of the model. The common features are extracted through the multi-step bidirectional long short-term memory network (bi-LSTM), and the output of the multi-step bidirectional long short-term memory network is used as the input of the multi-head attention mechanism. The multi-head attention mechanism extracts the common semantic information between samples by focusing on the common features between samples. As Figure 4 shown.

[0057] The multi-step bidirectional long short-term memory function is shown in the following formula:

[0058]

[0059] Among them, f and g represent neural networks, S represents the support set, K represents the number of iteration steps, represents the target feature.

[0060] The formula of the multi-head attention mechanism is shown as follows:

[0061] MultiHead(Q, K, V) = Concat(z 1 ,..., z h )w z ;

[0062] z i = Attention(QW i Q , KW i K , VW i V );

[0063] Among them, d k = d v = d model / h, where h is the number of heads of self-attention.

[0064] (3) Metric learning is to learn the distance or difference between data, which is used to effectively describe the similarity between samples. Extract features through a convolutional neural network and map the samples to a high-dimensional metric space, and measure the similarity between samples in the high-dimensional metric space; the measurement methods include Manhattan distance, Euclidean distance, Mahalanobis distance, Chebyshev distance, cosine distance, etc. In this implementation, the cosine distance is used to calculate the similarity between samples. When vectors a(x 11 , x 12 , x 13 ,..., x 1n ) and b(x 21 , x 22 , x 23 ,..., x 2n ) are in a vector space, then there is:

[0065]

[0066] Among them, a and b represent feature vectors.

[0067] Normalize the cosine distance, a = softmax(cosθ), and the predicted class of the input sample is:

[0068]

[0069] Among them, represents the query sample, x i represents the i-th class of support samples, y i represents the label of the i-th class of support samples, and k represents the number of support set categories.

[0070] Step 4: Train the few-shot recognition network model with the training set

[0071] Input the training set data into the few-shot recognition network model. As Figure 5 shown, in each training cycle, randomly select 3 classes from the dataset, and sample k samples from each class as the support set S and b samples as the query set Q. The training objective of the model is to maximize the probability that the support set S predicts the labels in the query set Q, and control the influence of easy and difficult samples on training. The target loss function is:

[0072] FL(p t ) = -α t (1 - p t ) γ log(p t );

[0073] Among them, α t is the weight parameter, p t is the predicted value, and γ is the adjustable factor. This loss function is used to solve the problem of imbalance between easy and difficult samples. (1 - p t ) γ is called the modulation coefficient. For samples with a relatively large probability, (1 - p t ) γ tends to 0, which can reduce the loss function value. For difficult samples with a relatively low true probability, (1 - p t ) γ has little impact on the loss function value. Use α t to increase the misclassification weight.

[0074] During the training process, verify with the validation set regularly. When the model reaches a certain accuracy, solidify the model parameters. The verification results are as Figure 6 shown. It can be seen from the figure that the verification accuracy of the differential experiment reaches 98%.

[0075] Step 5: Verify the few-shot recognition network model with the validation set

[0076] Regularly verify the network model with the validation set in the dataset constructed in Step 2. When the verification accuracy reaches the set value, solidify the model.

[0077] Step 6: Test the model

[0078] Input the radar echo signal into the solidified model, output the corresponding classification label, and finally obtain the category to which the target belongs.

[0079] After the small-sample recognition network model training is completed, the radar echo signal is input in real time, and the trained model parameters are loaded, and the target category can be output in real time.

[0080] By constructing a small-sample labeled dataset M((x i , y i ) ∈ M), input an unknown radar echo signal The output classification label is:

[0081]

[0082] Among them, f is the multi-head attention kernel function obtained by training in step S4, k is the number of categories of the dataset M, x i represents the labeled sample in the dataset M, and y i represents the label of x i .

[0083] The specific embodiments described above further elaborate on the purpose, technical solutions, and beneficial effects of the present invention. It should be understood that the above description is only the specific embodiments of the present invention and is not used to limit the protection scope of the present invention. Any modifications, equivalent replacements, improvements, 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 method for identifying radar echoes of low, slow and small targets based on small samples, characterized in that, it includes the following steps: Collect radar echo data and construct a radar echo database for low, slow and small targets; after preprocessing the collected radar echo data, construct a data set including a small training set, a validation set and a test set; the radar echo data includes the type of target and the coordinates, size, attitude angle and speed of the target; By constructing a small-sample labeled dataset M ((xi, yi) ∈ M), the classification label output by inputting an unknown radar echo signal is as follows: Among them, f is the trained multi-head attention kernel function, k is the number of categories of the dataset M, xi represents the labeled samples in the dataset M, and y i represents the label of x i . Construct a radar echo recognition network model, which includes a residual network feature extraction module, a memory enhancement module integrating a multi-head attention mechanism of a bidirectional long short-term memory network, and a similarity measurement module; The memory enhancement module integrating the multi-head attention mechanism of the bidirectional long short-term memory network organically combines the multi-step bidirectional long short-term memory network and the multi-head attention mechanism, extracts common features through the multi-step bidirectional long short-term memory network, and takes the output of the multi-step bidirectional long short-term memory network as the input of the multi-head attention mechanism; Among them, the multi-step bidirectional long short-term memory function is shown in the following formula: where f and g represent neural networks, S represents the support set, and K represents the number of iteration steps, represents the target feature; Use the data set to train, validate and test the radar echo recognition network model; The residual network feature extraction module is constructed by stacking 3 residual structures to form a 10-layer residual network for feature extraction; The formula of the multi-head attention mechanism is shown in the following formula: MultiHead(Q,K,V) = Concat(z 1 ,..., z h ) w z , z i = Attention(QW i Q , KW i K , VW i V ); Among them, d k = d v = d model / h, where h is the number of heads of self-attention; The similarity measurement module is used to learn the distance or difference between data and describe the similarity between samples; Extract features through a convolutional neural network and map samples to a high-dimensional metric space, and measure the similarity between samples in the high-dimensional metric space; the measurement means include Manhattan distance, Euclidean distance, Mahalanobis distance, Chebyshev distance and cosine distance; The goal of the convolutional neural network is to overcome the imbalance between easy and difficult samples, maximize the probability of predicting the test data set and minimize the loss function. The target loss function is: Among them, a t is the weight parameter, p t is the predicted value, and γ is an adjustable factor.

2. The method for identifying radar echoes of low, slow and small targets based on small samples according to claim 1, characterized in that, the preprocessing includes the following steps: First, delete invalid data, duplicate data, handle missing values and outliers in the constructed database; Secondly, construct a source task data set and a target task data set; and intercept the data, and the intercepted signal length is L; use the absolute amplitude detection method to detect the target in the cleaned data, set the threshold to ±0.75, and then intercept the target; among them, the intercepted length is L; Finally, extract features from the labeled samples; the feature extraction means include wavelet transform, frequency domain feature extraction and feature fusion.

Images