Radar emitter individual recognition method based on incremental neural network

Abstract

The present invention discloses a radar emitter individual identification method based on an incremental neural network. The method first obtains radar emitter data samples and extracts the time-frequency features and bi-spectrum features of the radar emitter data samples. Secondly, based on the ResNet1D network, combined with efficient channel attention ECA and self-organizing incremental learning neural network SOINN, a radar emitter individual identification network model is constructed. Finally, the time-frequency features and bi-spectrum features are combined into dual-channel samples, which are trained and tested through the radar emitter individual identification network model, and the radar emitter individual identification results are output to complete the radar emitter individual identification. The present invention realizes the screening of useful features, improves the learning ability of features, and then improves the accuracy of identification, and completes the accurate identification of radar emitter individuals.

Description

Technical Field

[0001] The present invention relates to the technical field of radar radiation source individual identification, and in particular to a radar radiation source individual identification method based on an incremental neural network. Background Art

[0002] The goal of radar emitter individual identification is to accurately distinguish different radar devices or different individuals of the same type of radar by analyzing the electromagnetic signals emitted by the radar. This technology can provide important support for situational awareness, target tracking and electronic countermeasures.

[0003] Traditional radar signal recognition methods usually rely on basic signal parameters such as pulse repetition frequency (PRF), pulse width, carrier frequency, etc. However, with the continuous development of radar technology, especially the widespread application of complex modulation methods, radar signals have become more complex, and traditional recognition methods based on manual feature extraction can no longer effectively cope with diverse and complex radar signal scenarios.

[0004] In order to solve this problem, radar signal processing technology based on deep learning has made significant progress in recent years. Through automatic feature extraction, deep neural networks (DNNs) can capture complex signal characteristics in multiple dimensions such as time domain, frequency domain, and energy domain, improving the accuracy of individual identification of radar emitters. In particular, convolutional neural networks (CNNs) perform well in processing one-dimensional time series signals. By extracting local features from signals through convolution operations, they greatly improve the efficiency and accuracy of signal processing. However, while deep neural networks have improved recognition performance, they also face some challenges, such as increased difficulty in model training, overfitting problems, and how to effectively capture global information in signals. Summary of the invention

[0005] The present invention proposes a radar emitter individual identification method based on incremental neural network, aiming to solve the problems of insufficient feature learning ability and low classification accuracy in the prior art. Existing radar emitter individual identification methods usually rely on a single feature for classification, but a single feature often cannot fully characterize the characteristics of the target. Therefore, the present invention combines time-frequency features with bispectral features, and uses multi-dimensional information for classification and identification, thereby improving the comprehensiveness and accuracy of identification.

[0006] The technical solution of the present invention is:

[0007] Step 1: Obtain radar radiation source data samples, and extract the time-frequency features and bispectral features of the radar radiation source data samples.

[0008] Step 2: Based on the ResNet1D network, combined with efficient channel attention (ECA) and self-organizing incremental learning neural network (SOINN), a radar emitter individual recognition network model is constructed to achieve deep feature extraction and fusion.

[0009] Step 3: Combine the time-frequency features and bispectral features obtained in step 1 into dual-channel samples, train and test them through the radar emitter individual identification network model constructed above, output the radar emitter individual identification results, and complete the radar emitter individual identification.

[0010] The present invention proposes a radar emitter individual recognition method based on incremental neural network. By extracting time-frequency features and bispectral features, weight allocation is performed according to ResNet1D and ECA enhancement to achieve feature fusion, and the constructed SOINN trained model is used for the classification and recognition of radar emitters.

[0011] Furthermore, in the above step 1, the time-frequency features and bispectral features of the extracted samples include:

[0012] The acquired radar radiation source data samples are preprocessed, and the preprocessing process includes pre-emphasis, framing and windowing.

[0013] After the preprocessing is completed, the preprocessed samples are subjected to short-time Fourier transform (STFT) and bispectral transform to obtain time-frequency features and bispectral features.

[0014] Perform STFT on the preprocessed samples to obtain the time-frequency characteristics of the signal:

[0015]

[0016] Where x[n] is the input signal, m is the time index, k is the frequency index, w[nm] is the time window function, usually a Hanning window, e -j2πkn / N is the Fourier basis function, and N is the number of Fourier transform points.

[0017] The bispectral features of the preprocessed samples are obtained by performing bispectral transformation again:

[0018]

[0019] The superscript * indicates complex conjugate, f1 and f2 are the two frequency components of the signal in the frequency domain, and are the axes of the two-dimensional plane of the bispectrum. The bispectral feature map is reduced to the dimension size of the time-frequency feature using the bicubic interpolation method, and the bispectral features and time-frequency features after dimensionality reduction are flattened to obtain two one-dimensional features.

[0020] Furthermore, in step 2, the feature vector first passes through two sequentially connected one-dimensional bottleneck residual blocks for feature extraction, and then enters the efficient channel attention module to assign feature weights, and then learns through the self-organizing incremental learning neural network, and finally realizes the classification and identification of radar radiation sources through the classifier constructed by the fully connected layer.

[0021] Furthermore, the structure of the one-dimensional bottleneck residual block is:

[0022] The structure of the one-dimensional bottleneck residual block consists of three convolutional units (each convolutional unit contains a cascaded convolutional layer, a batch normalization layer, and a ReLU activation layer), and a 3×3 convolutional layer is used to extract features. The residual connection uses a 1×1 convolutional layer to add the input of the convolutional unit to the features of the last batch normalization layer input of the convolutional unit.

[0023] The efficient channel attention layer automatically learns a set of weights after the input features pass through this network layer. The weights are used as thresholds to filter the features. The efficient channel attention is specifically the following process, assuming that the shape of the input tensor of ECA is C×H×W:

[0024] 1: Global pooling: For the input feature map X Cij (Input tensor), perform a global pooling operation on each channel to convert it into a scalar value. This process can be expressed by the following formula:

[0025]

[0026] Among them, f c It represents the scalar value obtained by the global pooling operation of the cth channel, and H and W represent the height and width of the feature map of the channel respectively.

[0027] 2: One-dimensional convolution operation: input feature map X Cij A one-dimensional convolution operation is performed on each channel to convert it into a scalar value.

[0028] 3: The feature map output by the one-dimensional convolution is normalized and mapped to the range of 0 to 1 to obtain the weight, which represents the importance weight of the channel.

[0029] 4: Use the obtained weights to weight the input feature map to obtain a new feature map. This process can be expressed by the following formula:

[0030] y Cij =W c ×X Cij

[0031] Among them, y Cij Represents the value of the i-th row and j-th column in the c-th channel in the new feature map.

[0032] 5: Superimpose the new feature maps of all channels in the channel direction: Add up the new feature maps of all channels in the channel direction to obtain the final output feature map of ECA.

[0033] Compared with the prior art, the present invention has the following advantages:

[0034] In order to solve the problems of poor anti-noise performance and weak feature learning ability of radar radiation source characteristic parameters, the proposed ResNet1D+ECA network gives higher weights to features with strong correlation, sets highly redundant features to zero, and fuses the thresholded features, thereby realizing the screening of useful features, improving the learning ability of features, and further improving the accuracy of recognition. SOINN can realize incremental learning, making the network more adaptable to the accurate recognition of new signals. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 is a flow chart of the method proposed by the present invention;

[0036] Figure 2 It is a schematic diagram of the overall framework of the designed network;

[0037] Figure 3 It is a one-dimensional bottleneck residual block structure diagram;

[0038] Figure 4 It is a graph showing the change of model accuracy with the number of iterations;

[0039] Figure 5 It is a confusion matrix diagram with an unknown sample. DETAILED DESCRIPTION

[0040] The present invention will be further described below in conjunction with the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solution of the present invention, and cannot be used to limit the protection scope of the present invention.

[0041] The present invention proposes a radar emitter individual identification method based on incremental neural network, aiming to solve the problems of insufficient feature learning ability and low classification accuracy in the prior art. Existing radar emitter individual identification methods usually rely on a single feature for classification, but a single feature often cannot fully characterize the characteristics of the target. Therefore, the present invention combines time-frequency features with bispectral features, and uses multi-dimensional information for classification and identification, thereby improving the comprehensiveness and accuracy of identification.

[0042] Reference Figure 1 The specific process of the present invention can be seen:

[0043] Step 1: Obtain radar radiation source data samples and extract the time-frequency characteristics and bispectral characteristics of the radar radiation source samples;

[0044] Step 2: If Figure 2 As shown in the figure, based on the ResNet1D network, combined with efficient channel attention (ECA) and self-organizing incremental learning neural network (SOINN), a radar emitter individual recognition network model is constructed to achieve deep feature extraction and fusion;

[0045] Step 3: Combine the two types of features obtained in step 1 into a dual-channel sample, train and test it through the radar emitter individual identification network model constructed above, and complete the radar emitter individual identification.

[0046] Furthermore, pre-emphasis compensates for the loss of high-frequency components of the signal, enhances the high-frequency components, and makes the spectrum of the signal smoother. The signal after pre-emphasis is expressed as

[0047] x'[n]=x[n]-α1x[n-1],α1=0.97

[0048] Where n is a discrete point.

[0049] Framing can divide the signal into several short-term signals, and the signal can be regarded as a stable process in the short period of time. Framing adopts the method of front and back overlap, and is generally performed at 1 / 4 or 1 / 2 of the sliding frame length. Windowing is to reduce spectrum leakage, reduce leakage frequency interference, and improve spectrum effects. In general, Hanning window is added, and its window function is expressed as:

[0050]

[0051] Among them, N is the number of sample points.

[0052] Suppose the signal is x'[n] and the window function is w[n]. The signal x[n] obtained after windowing is:

[0053] x[n]=x'[n]w[n],0≤n≤N-1

[0054] Step 1.2: After preprocessing, perform STFT on the signal to obtain the time-frequency diagram. The time-frequency diagram of the signal obtained by performing STFT on the signal is:

[0055]

[0056] Among them, k is the frequency index and m is the time index.

[0057] Perform bispectral transformation on the signal again to obtain the bispectrum:

[0058]

[0059] The bispectrum is reduced to the dimension of the time-frequency graph using the bicubic interpolation method, and the bispectrum and time-frequency graph after dimension reduction are directly flattened to obtain two one-dimensional features.

[0060] Furthermore, the feature vector in step 2 first passes through two sequentially connected one-dimensional bottleneck residual blocks for feature extraction, then enters the efficient channel attention module to assign feature weights, and then learns through the self-organizing incremental learning neural network, and finally realizes the classification and recognition of radar emitters through the fully connected layer, as shown in Figure 2 shown.

[0061] Table 1 Parameters of one-dimensional bottleneck residual block

[0062] Tiers Number of output channels parameter Conv1_1 4 (2，4，1，1) Conv1_2 4 (4，4，3，1) Conv1_3 8 (4，8，1，1) Conv2_1 16 (8，16，1，1) Conv2_1 16 (16，16，3，1) Conv2_3 32 (16，32，1，1)

[0063] Step 2.1: The one-dimensional bottleneck residual block specifically goes through three convolutional layers. The overall process is:

[0064] Set the input feature dimension to 1×2×N. After Conv1_1, the output size is 1×4×N. After Conv1_2, the output size remains unchanged. After Conv1_3, the output size is 1×8×N. At this time, it is input to the second residual block. After Conv2_1, the output size is 1×16×N. After Conv2_2, the output size remains unchanged. After Conv2_3, the output size is 1×32×N. The one-dimensional bottleneck residual block structure is shown in the figure below: Figure 3 shown.

[0065] Step 2.2: Input the obtained features into ECA, perform a global pooling operation on each channel, and convert it into a scalar value of size 1×32×1. Perform a one-dimensional convolution operation on each channel and convert it into a scalar value of size 1×32×1. Map the scalar value to a range between 0 and 1 through the sigmoid function: Map the scalar value obtained in step 2 to a range between 0 and 1, indicating the importance weight of the channel. Perform a weighted average of the input features, size 1×32×N: The weights obtained in step 3 are used to perform a weighted average of the features of the channel to obtain a new feature. Add up the new features of all channels in the channel direction: Add up the new features of all channels in the channel direction to obtain the final output feature size of 1×32×N.

[0066] Step 2.3: Flatten the final output feature into a one-dimensional vector 1×N, input it into the SOINN network, find the closest neuron, and update the neuron if the distance between the nearest neuron and the input feature is below a certain threshold, otherwise create a new neuron. This step does not change the dimension of the input vector. The output of SOINN is a category or neuron index, indicating the classification result of the input signal.

[0067] Furthermore, the training and recognition process of step 3 is:

[0068] The extracted features were put into the constructed ResNet1D and ECA enhanced incremental neural networks and the standard ResNet1D network to obtain the radar emitter category prediction results and compare the recognition accuracy of the two. The maximum number of nodes in SOINN was set to 200. The threshold was set to 0.1 to decide when to insert a new node. The initial learning rate was set to 0.01 to optimize the network prediction results. After multiple rounds of iterations, the final recognition model was obtained. Figure 4 The accuracy of the model changes with the number of iterations. Then put the test features into the trained recognition model, perform target testing and recognition, and generate a confusion matrix as shown in Figure 5 shown.

[0069] The specific embodiments are:

[0070] 1: Get the data set, specifically:

[0071] Radar signals are collected through different types of radar equipment, and radar equipment is divided into categories: Small category A: pulse radar, short-wave radar, short-range navigation radar. Smaller category B: mobile target indication (MTI) radar, vehicle-mounted radar. Larger category C: medium-range surveillance radar, weather radar. Larger category D: long-range detection radar, high-power pulse radar.

[0072] 2: Data preprocessing, specifically:

[0073] In order to avoid the overfitting of the model due to the small number of intermediate frequency data samples in the original data set and the imbalance of the number of samples of different categories, all intermediate frequency data are divided into 3ms segments in data preprocessing and converted into corresponding time-frequency diagrams and bispectral diagrams. The specific processing parameters are as follows: First, the original audio is downsampled, the sampling frequency is set to 1000Hz, the window length is 256ms, and the window length is 1 / 4 of each sliding window; the number of Fourier transform points is 512. After processing, it is ensured that each category contains 266 samples, and the total number of samples reaches 1330. The sample set is divided into a training set and a validation set in a ratio of 8:2, where the validation set contains radar category data that does not appear in the training set.

[0074] 3: Network parameter settings, specifically:

[0075] Set the input feature dimension to 1×2×N. After Conv1_1, the output size is 1×4×N. After Conv1_2, the output size remains unchanged. After Conv1_3, the output size is 1×8×N. At this time, it is input to the second residual block. After Conv2_1, the output size is 1×16×N. After Conv2_2, the output size remains unchanged. After Conv2_3, the output size is 1×32×N. The input dimension of the ECA module is 32. Set the maximum number of nodes of SOINN to 200. Set the threshold to 0.1 to decide when to insert a new node.

[0076] 4: Training parameter settings, specifically:

[0077] For the ResNet1D+ECA module, the training learning rate is set to 0.001, the number of rounds is 100, the loss function uses the cross entropy loss function, and the training optimizer is Adam with a batch size of 32. For SOINN, only the initial learning rate needs to be set to 0.1.

[0078] 5: In order to evaluate the model, the accuracy and F1-score are used as evaluation indicators for the designed model and the comparison model. The F1-score is obtained by weighting the precision and recall. The specific calculation formula is as follows:

[0079] Accuracy:

[0080]

[0081] Recall:

[0082]

[0083] Accuracy:

[0084]

[0085] F1-score:

[0086]

[0087] Among them, TP is correctly predicted as a positive example, TN is correctly predicted as a negative example, FP is incorrectly predicted as a positive example, and FN is incorrectly predicted as a negative example.

[0088] 6: Analysis of recognition results, specifically:

[0089] (1) First, compare the recognition effects of time-frequency features and bispectral features on the model; then add time-frequency features and bispectral features at the same time to observe the recognition effect of the fusion method.

[0090] As shown in Table 2, after fusing time-frequency features and bispectral features, the recognition accuracy and F1 score can be improved by about 1%, indicating that multi-dimensional features can improve recognition accuracy.

[0091] Table 2. Results of feature fusion scheme (unit: %)

[0092] Feature Scheme Accuracy F1-score Time-frequency 95.14 94.64 Double spectrum 95.32 94.27 Time-frequency + bispectrum (the present invention) 96.65 95.63

[0093] (2) The designed network model is compared with the basic residual model, and a network model without the last SOINN module is set as the control group.

[0094] As shown in Table 3, after adding ECA, the recognition accuracy is improved by about 1.5%, indicating that ECA reduces the redundant noise in the data set and is excellent in the fusion of useful features. On this basis, adding SOINN improves the recognition accuracy by 4%, indicating that SOINN can recognize data samples that have not appeared in the training set.

[0095] Table 3 Model design results (unit: %)

[0096] Model design plan Accuracy F1-score ResNet1D 90.84 90.67 ResNet1D+ECA 92.36 91.82 ResNet1D+ECA+SOINN (the present invention) 96.38 95.94

[0097] In summary, the radar emitter individual identification method based on incremental neural network proposed in the present invention has higher identification accuracy than other models both in terms of features and network design.

[0098] The above is only a preferred embodiment of the present invention and does not limit the present invention in any way. Any simple modification, change and equivalent change made to the above embodiment according to the technical essence of the invention still falls within the protection scope of the technical solution of the present invention.

Claims

1. A radar emitter individual identification method based on incremental neural network, characterized in that: The following steps are involved: Step 1: Obtain radar radiation source data samples, and extract time-frequency features and bispectral features of the radar radiation source data samples; Step 2: Based on the ResNet1D network, combined with the efficient channel attention ECA and the self-organizing incremental learning neural network SOINN, a radar emitter individual recognition network model is constructed; The radar emitter individual identification network model is implemented as follows: the feature vector is first extracted by two sequentially connected one-dimensional bottleneck residual blocks, then input into the efficient channel attention module to assign feature weights, and then learned by the self-organizing incremental learning neural network, and finally the radar emitter is classified and identified by the classifier constructed by the fully connected layer; The efficient channel attention ECA is to automatically learn a set of weights after the input features pass through the network layer, and the weights are used as thresholds to filter the features; Efficient channel attention specifically includes the following process, assuming that the shape of the input tensor of ECA is C×H×W: For the input feature map X Cij , perform a global pooling operation on each channel and convert it into a scalar value; For the input feature map X Cij One-dimensional convolution operation is performed on each channel; The feature map output by the one-dimensional convolution is normalized and mapped to the range of 0 to 1 to obtain the weight; The obtained weights are used to input feature map X Cij Perform weighting to obtain a new feature map; The new feature maps of all channels are superimposed in the channel direction to obtain the final output feature map of ECA; Step 3: Combine the time-frequency features and bispectral features into dual-channel samples, train and test them through the radar emitter individual recognition network model, and output the radar emitter individual recognition results.

2. The radar emitter individual identification method based on incremental neural network according to claim 1 is characterized in that: The specific process of extracting the time-frequency characteristics and bispectral characteristics of the radar radiation source data sample is as follows: Preprocessing the acquired radar radiation source data samples, including pre-emphasis, framing and windowing; After the preprocessing is completed, the preprocessed samples are subjected to short-time Fourier transform (STFT) and bispectral transform to obtain time-frequency features and bispectral features; the details are as follows: Perform STFT on the preprocessed samples to obtain the time-frequency characteristics of the signal: Where x[n] is the input signal, m is the time index, k is the frequency index, w[nm] is the time window function, and e[nm] is the frequency index. -j2πkn / N is the Fourier basis function, N is the number of Fourier transform points; The bispectral features obtained by performing bispectral transformation on the preprocessed samples are: Wherein, the superscript * indicates complex conjugate, f1 and f2 are the two frequency components of the signal in the frequency domain, and are the axes of the two-dimensional plane of the bispectrum; The bispectral features are reduced to the dimension size of the time-frequency features using the bicubic interpolation method, and the reduced bispectral features and time-frequency features are flattened to obtain two one-dimensional features.

3. The radar emitter individual identification method based on incremental neural network according to claim 2 is characterized in that: The structure of the one-dimensional bottleneck residual block is as follows: The structure of the one-dimensional bottleneck residual block consists of three convolutional units. Each convolutional unit contains a cascaded convolutional layer, a batch normalization layer, and a ReLU activation layer. The residual connection uses a 1×1 convolutional layer to add the input of the convolutional unit to the features of the last batch normalization layer input of the convolutional unit.

Images