A continuous learning method for radar HRRP based on generative adversarial network

Abstract

The present invention discloses a radar HRRP continuous learning method based on a generative adversarial network, comprising the following steps: S1, using measured data including three types of aircraft, namely, a medium-sized propeller aircraft An‑26, a small jet aircraft Cessna, and a large jet aircraft Yark‑42, the radar operates in the C band, has a signal bandwidth of 400 MHz, and a pulse repetition frequency of 400 Hz; S2, processing intensity sensitivity and translation sensitivity in a preprocessing process; S3, using three different HRRP incremental learning task settings under the background of radar HRRP automatic target recognition; S4, training the HRRP data processed by S2 under the three settings according to a continuous learning related method, and redesigning the decoder for the pseudo-replay DGR method to a conditional autoencoder adversarial network CVAEGAN, wherein the CVAEGAN includes an encoder, a decoder, a discriminator, and a classifier; S5, training the HRRP data processed by S2 using a DGR method of the CVAEGAN network.

Description

Technical Field

[0001] The present invention belongs to the field of radar technology, and in particular relates to a radar HRRP continuous learning method based on a generative adversarial network. Background Art

[0002] The range resolution of a high-resolution broadband radar is much smaller than the target size. Its echo is also known as the target's one-dimensional high-resolution range profile (HRRP). The HRRP not only contains structural information such as the target's radial size and the distribution of scattering points, but is also relatively simple to acquire and process, and convenient to store, making it highly suitable for engineering applications. Consequently, target recognition methods based on HRRP have garnered widespread attention in the field of radar automatic target recognition.

[0003] Previous research on HRRP object recognition has mostly focused on static, offline tasks. These tasks first involve collecting training HRRP data and building a sample library. Then, they train object recognition models based on statistical or deep learning models. Finally, these models are deployed offline on hardware devices for real-world recognition tasks. While these approaches have achieved promising results, they are unable to scale online to acquire recognition capabilities for new HRRP samples, hindering their application in diverse, real-world scenarios. Summary of the Invention

[0004] In view of the above technical problems, the present invention provides a radar HRRP continuous learning method based on a generative adversarial network.

[0005] In order to solve the above technical problems, the present invention adopts the following technical solutions:

[0006] A radar HRRP continuous learning method based on a generative adversarial network includes the following steps:

[0007] S1, the measured data used includes three types of aircraft, namely the medium-sized propeller aircraft An-26, the small jet aircraft Cessna and the large jet aircraft Yark-42. The radar operates in the C band, the signal bandwidth is 400MHz, and the pulse repetition frequency is 400Hz. Each HRRP sample in the data set contains 256 range units. The 2nd and 5th segments of the An-26 aircraft, the 6th and 7th segments of the Cessna aircraft, and the 5th and 6th segments of the Yark-42 aircraft are used as training samples, and the remaining segments are used as test samples. The training data are extracted using the equal interval sampling method to make the number of training samples the same as the simulation data, and a certain amount of Gaussian white noise is added to make its signal-to-noise ratio 25dB. The dataset of 9 types of aircraft used is the electromagnetic simulation HRRP dataset generated by FECO software based on the turntable model. Each type of aircraft dataset used for training covers 360 degrees in the azimuth dimension and contains 1600 training samples. Each HRRP sample contains 256 range units. The target type, azimuth dimension coverage angle domain and number of samples of the simulated aircraft dataset used for testing are consistent with those of the aircraft dataset used for training. There is only a difference of about 10 degrees between the pitch angles. In order to make the simulation data more realistic, a certain amount of Gaussian white noise is added during the simulation to make the signal-to-noise ratio of the simulation dataset 25dB.

[0008] S2, intensity sensitivity and translation sensitivity are processed during preprocessing;

[0009] S3, in the context of radar HRRP automatic target recognition, uses three different HRRP incremental learning task settings;

[0010] S4, according to the continuous learning related method, trains the HRRP data processed by S2 under three settings, and redesigns the decoder for the pseudo-replay DGR method as the conditional autoencoder adversarial network CVAEGAN, where CVAEGAN includes an encoder, a decoder, a discriminator and a classifier;

[0011] S5, uses the DGR method of the CVAEGAN network to train the HRRP data processed by S2.

[0012] Preferably, the S2 further comprises:

[0013] S201, perform L2 normalization processing on the original HRRP echo, and the original HRRP data is represented by X raw =[x1, x2, ..., x M ], where M represents the total number of HRRP distance units, and the normalized x normalization Expressed as:

[0014]

[0015] where xi represents the intensity of the i-th distance unit;

[0016] S202, using the center of gravity alignment method to eliminate translation sensitivity. In the center of gravity alignment method, the center of gravity position of the data is first calculated, and then the center of gravity of the HRRP data is shifted left and right to bring it closer to the center of the data. The calculation process of the center of gravity g is expressed as:

[0017]

[0018] Preferably, the three different HRRP incremental learning tasks in S3 include task incremental learning, domain incremental learning and class incremental learning.

[0019] The present invention has the following beneficial effects:

[0020] (1) It improves the catastrophic forgetting problem that existed in the offline training model in the past, allowing the radar to update its parameters in real time based on new data on the battlefield, thus expanding its practical application range;

[0021] (2) The proposed model not only has certain data privacy and fixed model capacity characteristics, but also has better recognition effect than the model based on regularization method. DETAILED DESCRIPTION

[0022] The following is a clear and complete description of the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making any creative efforts are within the scope of protection of the present invention.

[0023] In actual battlefield environments, new HRRP data for non-cooperative targets may be collected. If the parameters of a previously deployed offline model are directly updated using these new samples, the classification knowledge acquired through offline learning will be catastrophically forgotten. Alleviating this catastrophic forgetting problem and ensuring the model's scalability to new samples—in other words, retaining the target recognition capabilities acquired through previous training while also being able to recognize these new targets—has become a pressing engineering challenge.

[0024] A radar HRRP continuous learning method based on a generative adversarial network according to an embodiment of the present invention includes the following steps:

[0025] S1. Due to the lack of measured data, a combination of measured data and simulation data was used to verify the effectiveness of the method of the embodiment of the present invention. The measured data used included three types of aircraft. The radar operates in the C band, with a signal bandwidth of 400 MHz and a pulse repetition frequency of 400 Hz. The three aircraft are the medium-sized propeller aircraft "An-26", the small jet aircraft "Cessna", and the large jet aircraft "Yark-42". Each HRRP sample in the dataset contains 256 range units. To ensure that the azimuth and angular domains of the training data cover the azimuth and angular domains of the test data, segments 2 and 5 of the An-26 aircraft, segments 6 and 7 of the Cessna aircraft, and segments 5 and 6 of the Yark-42 aircraft are used as training samples, and the remaining segments are used as test samples. In addition, to avoid class imbalance among the samples, the training data is sampled using equal interval sampling to ensure the same number of training samples as the simulation data. A certain amount of Gaussian white noise is added to ensure a signal-to-noise ratio of 25 dB. The datasets used for the nine aircraft types are electromagnetic simulation HRRP datasets generated by FECO software based on a turntable model. Each aircraft type used for training covers a 360-degree angular domain in the azimuth dimension and contains 1,600 training samples, with each HRRP sample consisting of 256 range cells. The simulated aircraft datasets used for testing match the training datasets in target type, azimuth coverage, and sample number, with the only difference being approximately 10 degrees in pitch angle. To enhance the realism of the simulated data, a certain amount of Gaussian white noise was added during the simulation to achieve a signal-to-noise ratio of 25dB.

[0026] S2, since the collected original HRRP data contains intensity sensitivity and translation sensitivity, these two sensitivities will be processed during the preprocessing process to eliminate their instability effects on the back-end deep neural network model;

[0027] S3, in the context of radar HRRP automatic target recognition, in order to make the comparison between various incremental learning methods more reflective of actual performance, three different HRRP incremental learning task settings are used;

[0028] S4, according to the continuous learning related method, trains the HRRP data in S2 under three settings, and redesigns the decoder for the pseudo-replay DGR method to generate the adversarial network CVAEGAN, where CVAEGAN includes encoder, decoder, discriminator, and classifier;

[0029] S5, by using the DGR method of the CVAEGAN network to train the HRRP data in S2, the final test results are superior to other methods.

[0030] In one embodiment of the present invention, S2 further includes:

[0031] S201, in order to eliminate the influence of intensity sensitivity, the original HRRP echo is subjected to L2 normalization. If the original HRRP data can be expressed as x raw =[x1, x2, ..., x M ], where M represents the total number of HRRP distance units, and the normalized x normalization It can be expressed as:

[0032]

[0033] where x i Represents the intensity of the i-th distance unit.

[0034] S202, using the center of gravity alignment method to eliminate translation sensitivity. In the center of gravity alignment method, the center of gravity position of the data is first calculated, and then the center of gravity of the HRRP data is shifted left and right to make it close to the center of the data. The calculation process of the center of gravity g can be expressed as:

[0035]

[0036] In one embodiment of the present invention, S3 further includes:

[0037] S301, first divide the data set into several different tasks. The first task setting of HRRP incremental learning is called task incremental learning, represented by Task-IL. In Task-IL, the task number to which the test sample belongs is known, and the model only needs to point out which task number the tested HRRP sample belongs to. The second task setting of HRRP incremental learning is called domain incremental learning, represented by domain-IL. Under this setting, we only focus on which category of targets in the incremental learning task the test HRRP sample corresponds to. The third setting of HRRP incremental learning tasks is called class incremental learning, represented by class-IL. After given a test HRRP sample, the model needs to output the target corresponding to the HRRP sample.

[0038] In one embodiment of the present invention, S4 further includes:

[0039] S401, the purpose of the encoder is to obtain the mean and variance of the random variable z used for the decoder to generate HRRP. The label y corresponding to the HRRP data x is input into the encoder network together with the data as a condition, that is, the conditional label y is encoded through one-hot encoding to obtain an N-dimensional vector x_hat, and then directly concatenated with x to obtain x_input[x,x_hat] as a new vector as input. This input is encoded by a combination of two fully connected layers and ReLU activation functions, and finally the obtained encoding is passed through different fully connected layers to obtain the mean and variance of the decoder input z. The formula for the data passing through the fully connected layer and ReLu can be expressed as:

[0040] y FC =f(Wx input +b)

[0041] Where W represents the weight matrix of the fully connected layer, b represents the bias of the fully connected layer, and the function f(·) represents the ReLU activation function.

[0042] f(x)=max(0,x)

[0043] In S402, the decoder generates an HRRP sample corresponding to the category using a random variable and the conditional category label y. The encoder and decoder form an encoder-decoder pair, so their designs are relatively similar. Similar to the encoder, this input is passed through a combination of two fully connected layers and a ReLU activation function to extract features. In the final combination used to generate HRRP samples, the activation function is replaced with a sigmoid activation function, which keeps the output value between 0 and 1.

[0044]

[0045] Where exp(·) represents the exponential operation.

[0046] S403: Use the discriminator to distinguish between "real" and "fake" samples. To reduce the risk of model collapse in the CVAE-GAN model and make the HRRP samples generated by the decoder more realistic, we use a more complex discriminator.

[0047] S404 aims to ensure that the reconstructed data, after conditional encoding and decoding, still has the same categorizability as the real data, thereby ensuring that the data generated by the generator meets its conditions (labels). The classifier network also uses two fully connected layers with ReLU activation functions to extract features, and finally uses a fully connected layer and softmax to obtain the discrimination result. Unlike the output discrimination result in GAN (which only distinguishes true from false), the discrimination range in the final classifier is determined by the three task settings.

[0048] We assume that the input before softmax is represented by input. If the total number of targets in the training set is C, the output after softmax can be represented as:

[0049]

[0050] Here, exp(·) represents exponential operation, input(i) and output(i) refer to the i-th element in the vectors input and output, respectively. Output(i) is also the maximum a posteriori probability P(i|x), which represents the probability that x is classified as label i.

[0051] In one embodiment of the present invention, S5 further includes:

[0052] In S501, the HRRP sample is first transformed in dimension to pass through the SE-Block. In our experiments, we set the number of channels to 256 so that this module can focus more on the more important distance units in the HRRP sample. We pass x through a layer composed of two fully connected modules to obtain the weight adjustment value c. The final result x_se can be expressed as:

[0053] x SE (i)=c(i)×x(i)

[0054] Where x is the input of SE-block, c is the weight adjustment result, x SE is the output of SE-block, and the i in the brackets represents the i-th element in the corresponding vector.

[0055] S502: After adjusting the importance of each layer, x_se is sent to a module consisting of two layers of Transformers. The two layers of Transformers represent two identical encoders and two identical decoders.

[0056] The overall input is x_se, and the 256 HRRP distance units of x_se are fed into the self-attention module. This layer uses not only the information of a distance unit itself but also the information of other distance units when processing a distance unit. The z processed by self-attention is as follows:

[0057]

[0058] Where Q, K and V are the input x_se and the three weight matrices W Q , W K and W VThe result of multiplication represents the matrix of Query, Key, and Value respectively. k Represents the length of the key vector, and the superscript T of K indicates that it is a transposed matrix.

[0059] After passing the self-attention module, we perform Add&Normalize operations:

[0060] z=LayerMorm(x_se+z)

[0061] Among them, LayerNorm represents layer normalization, and x_se+z represents residual connection, i.e. Add.

[0062] Then the residual is normalized through the feedforward neural network. The results after the feedforward neural network are as follows:

[0063] z=FFNN(z)=max(0,zW1+b1)W2+b2

[0064] The feedforward neural network consists of two layers of fully connected networks and activation functions. The activation function of the first layer is ReLU, and the second layer is a linear activation function. That is, W1, b1 and W2, b2 represent the weight matrix and bias of the two fully connected layers, respectively.

[0065] The result of this processing is then processed once through the same encoder, and finally the result obtained after two encoders is a set of attention vectors K and V.

[0066] Most of the components of the decoder are similar to those of the encoder, and its input is also x_se, which is fed into the self-attention. The difference is that the self-attention in the decoder only allows attention to the information before the current distance unit instead of all.

[0067] The principle of the Encoder-Decoder Attention layer is similar to the multi-head attention mechanism. The difference is that it uses the output of the previous layer to construct the Query matrix, while the Key matrix and Value matrix come from the final output of the encoder.

[0068]

[0069] Among them, Q comes from the previous layer of the decoder, while K and V come from the encoder.

[0070] It should be understood that the exemplary embodiments described herein are illustrative and not restrictive. Although one or more embodiments of the present invention have been described, it should be understood by those skilled in the art that various changes in form and detail may be made without departing from the spirit and scope of the present invention as defined by the appended claims.

Claims

1. A radar HRRP continuous learning method based on generative adversarial network, characterized by: The following steps are involved: S1, the measured data used includes three types of aircraft, namely the medium-sized propeller aircraft An-26, the small jet aircraft Cessna and the large jet aircraft Yark-42. The radar operates in the C band, the signal bandwidth is 400MHz, and the pulse repetition frequency is 400Hz. Each HRRP sample in the data set contains 256 range units. The 2nd and 5th segments of the An-26 aircraft, the 6th and 7th segments of the Cessna aircraft, and the 5th and 6th segments of the Yark-42 aircraft are used as training samples, and the remaining segments are used as test samples. The training data are extracted using the equal interval sampling method to make the number of training samples the same as the simulation data, and a certain amount of Gaussian white noise is added to make its signal-to-noise ratio 25dB. The dataset of 9 types of aircraft used is the electromagnetic simulation HRRP dataset generated by FECO software based on the turntable model. Each type of aircraft dataset used for training covers 360 degrees in the azimuth dimension and contains 1600 training samples. Each HRRP sample contains 256 range units. The target type, azimuth dimension coverage angle domain and number of samples of the simulated aircraft dataset used for testing are consistent with those of the aircraft dataset used for training. There is only a difference of about 10 degrees between the pitch angles. In order to make the simulation data more realistic, a certain amount of Gaussian white noise is added during the simulation to make the signal-to-noise ratio of the simulation dataset 25dB. S2, intensity sensitivity and translation sensitivity are processed during preprocessing; S3, in the context of radar HRRP automatic target recognition, uses three different HRRP incremental learning task settings; S4, according to the continuous learning related method, trains the HRRP data processed by S2 under three settings, and redesigns the decoder for the pseudo-replay DGR method. The decoder is specifically the conditional autoencoder adversarial network CVAEGAN, where CVAEGAN includes an encoder, a decoder, a discriminator, and a classifier; S5, uses the DGR method of the CVAEGAN network to train the HRRP data processed by S2.

2. The radar HRRP continuous learning method based on a generative adversarial network according to claim 1, characterized in that: Said S2 further comprises: S201, perform L2 normalization processing on the original HRRP echo, and the original HRRP data is represented as x raw =[x1,x2,...,x M ], where M represents the total number of HRRP distance units, and the normalized x normalization Expressed as: where x i represents the intensity of the i-th distance unit; S202, using the center of gravity alignment method to eliminate translation sensitivity. In the center of gravity alignment method, the center of gravity position of the data is first calculated, and then the center of gravity of the HRRP data is shifted left and right to bring it closer to the center of the data. The calculation process of the center of gravity g is expressed as:

3. The radar HRRP continuous learning method based on a generative adversarial network according to claim 1, characterized in that: The three different HRRP incremental learning tasks in S3 include task incremental learning, domain incremental learning, and class incremental learning.

Images