Radar signal pattern recognition method based on multi-modal feature fusion

Abstract

The invention discloses a radar signal pattern recognition method based on multi-modal feature fusion, and the method comprises the steps: obtaining in-pulse IQ sampling data and full-pulse PDW data of a radar signal, and carrying out the time-frequency transformation of the IQ sampling data, and obtaining a time-frequency image; respectively preprocessing the time-frequency image and the full-pulse PDW data; performing feature extraction on the preprocessed time-frequency image and the full-pulse PDW data to obtain intra-pulse features and inter-pulse features; carrying out multi-modal feature adaptive fusion on the intra-pulse features and the inter-pulse features; and the fused features are used as classifier input data for classification and identification, and radar signal pattern identification is realized. The method shows high recognition accuracy under different signal-to-noise ratio conditions, can still maintain good performance especially under the condition of low signal-to-noise ratio, and can effectively improve the recognition accuracy and robustness.

Description

Technical Field

[0001] The present invention belongs to the field of signal recognition, and relates to a radar signal pattern recognition technology, and specifically to a radar signal pattern recognition method based on multi-modal feature fusion. Background Art

[0002] With the development of phased array radar and digital array radar technology, the performance and functions of modern radar have been significantly improved. Phased array radar can quickly adjust its beam direction to achieve multi-target tracking and high-precision detection. Digital array radar improves the radar's anti-interference ability and resolution through digital signal processing technology. The functions of airborne radar are more diversified.

[0003] The radar's operating mode often switches rapidly according to mission requirements. For example, in air alert mode, the radar needs to cover a wide scanning range to detect potential threats in a timely manner; in target tracking mode, the radar needs to focus the beam to accurately track a specific target; in electronic countermeasure mode, the radar may need to adjust its signal parameters to avoid detection and interference. This variety of operating modes makes the identification and analysis of airborne radars more complicated.

[0004] In addition, the diversity and complexity of modern radar signal patterns also increase the difficulty of identification. Radar signals can use a variety of modulation methods such as linear frequency modulation (LFM), binary phase shift keying (BPSK), nonlinear frequency modulation (NLFM), etc. Different modulation methods have different signal characteristics and uses, and play different roles in different scenarios.

[0005] Since radar frequently switches working modes and signal modulation methods in different scenarios, accurate identification and analysis requires the combination of multiple advanced technologies. In order to address the problem of insufficient features of a single data source, combining working modes and signal modulation methods can provide more comprehensive target information.

[0006] At present, radar signal recognition methods are mainly divided into inter-pulse modulation feature signal recognition and intra-pulse modulation feature recognition. Inter-pulse modulation feature sorting algorithms include template matching method and PRI-based sorting algorithm. The template matching method is relatively simple. It is based on the existing feature database to compare the received signal pulse description words, which is suitable for relatively simple scenarios. As the electromagnetic environment becomes increasingly complex, researchers have begun to use the timing pattern of radar pulse parameters to characterize radar signals, and combined with deep learning technology, it provides excellent performance for the recognition of radar signal working modes. However, in the case of overlapping radar signal parameters, the performance of this method will still be affected. Summary of the invention

[0007] Purpose of the invention: To address the problems of single feature form and insufficient feature expression shown by radar pattern recognition using only intra-pulse data or inter-pulse pulse descriptors, a radar signal pattern recognition method based on multimodal feature fusion is provided. The method shows high recognition accuracy under different signal-to-noise ratio conditions, especially under low signal-to-noise ratio conditions, and can still maintain good performance, which can effectively improve the accuracy and robustness of recognition.

[0008] Technical solution: To achieve the above object, the present invention provides a radar signal pattern recognition method based on multimodal feature fusion, comprising the following steps:

[0009] S1: Obtain the IQ sampling data and full pulse PDW data of the radar signal pulse, and perform time-frequency transformation on the IQ sampling data to obtain a time-frequency image;

[0010] S2: preprocessing the time-frequency image and full-pulse PDW data respectively;

[0011] S3: extracting features from the preprocessed time-frequency image and full-pulse PDW data respectively to obtain intra-pulse features and inter-pulse features;

[0012] S4: Adaptive fusion of multimodal features for intra-pulse features and inter-pulse features;

[0013] S5: Use the fused features as classifier input data for classification and recognition to achieve radar signal pattern recognition.

[0014] Furthermore, the preprocessing method of the time-frequency image in step S2 is: based on the brightness level of the image pixel, establish the corresponding relationship between the brightness Y and the three color channels of R, G, and B: Y = 0.2989R + 0.5870G + 0.1140B, perform denoising on the time-frequency image, filter out the noise of the grayscale processed image using a smoothing filter method, and convert the denoised image into a grayscale image using a weighted average method. The core of this method is to eliminate the hue and saturation information of the image, express the grayscale value of the image as a brightness value, and realize the grayscale processing of the image.

[0015] Furthermore, the preprocessing method of the full-pulse PDW data in step S2 is: expanding the full-pulse PDW data of the radar signal, that is, expanding the PDW data into a one-dimensional feature vector.

[0016] Furthermore, in step S3, feature extraction is performed on the preprocessed time-frequency image through a high sampling rate intra-pulse IQ data feature extraction network to obtain intra-pulse features; and feature extraction is performed on the preprocessed full-pulse PDW data through a low sampling rate PDW data feature extraction network to obtain inter-pulse features.

[0017] Furthermore, the high sampling rate intra-pulse IQ data feature extraction network is a 2DResNet network, which extracts features through 1 2D convolution layer and 3 residual modules. The 3 residual modules have the same architecture. Each residual module contains 2 2D convolution layers and corresponding batch normalization layers and activation functions. Residual connections are added between modules to adapt to deeper networks; each convolution layer is equipped with 1 batch normalization layer and 1 ReLU activation function for adaptively adjusting the changing gradient; after the last residual module, a global average pooling is used, which can process input data of different lengths so that the output features are consistent, to extract important inter-pulse features of each channel, and obtain the image feature matrix as the intra-pulse features.

[0018] Furthermore, the low sampling rate PDW data feature extraction network is a 1DResNet network, which extracts features through 1 1D convolution layer, 3 residual modules and 1 fully connected layer. The 3 residual modules have the same architecture. Each residual module contains 2 1D convolution layers and corresponding batch normalization layers and activation functions. Residual connections are added between modules to adapt to deeper networks; each convolution layer is equipped with 1 batch normalization layer and 1 ReLU activation function for adaptively adjusting the changing gradient; after the last residual module, a global average pooling is used, which can process input data of different lengths so that the output features are consistent, to extract important inter-pulse features of each channel, and obtain the image feature matrix as the inter-pulse features.

[0019] Furthermore, the adaptive fusion method in step S4 is:

[0020] A1: Adaptive pooling operations are applied to inter-pulse features and intra-pulse features respectively to obtain the pooling results of inter-pulse features and intra-pulse features;

[0021] A2: Fuse the pooling results of inter-pulse features and intra-pulse features to obtain a fused feature vector.

[0022] Furthermore, the step A2 is specifically as follows:

[0023] Define two learnable parameters α and β to weight the feature vectors of 1D ResNet and 2D ResNet. The fused feature vector is expressed as:

[0024] Fused Feature=α×1D? Feature+β×2D? Feature (1)

[0025] To ensure that the weights sum to 1, the Softmax function is used to normalize the weights:

[0026] α,β=softmax([w1,w2]) (2)

[0027] Where w1 and w2 are two learnable parameters.

[0028] Furthermore, the classifier in step S5 is a trained deep metric network, and the fused feature vector is input into the deep metric network to learn the feature similarity metric space and output the final recognition result.

[0029] The present invention explores a method of combining intra-pulse and inter-pulse features, making full use of the multi-modal characteristics of radar signal patterns, aiming to add one more dimension of features for joint judgment to improve the accuracy of signal recognition when it is difficult to distinguish overlapping parameters of different radar signals.

[0030] Beneficial effects: Compared with the prior art, the present invention uses PDW data and time-frequency grayscale images as different information sources to input into the ResNet network, and can obtain two different types of feature representations; uses adaptive pooling for feature fusion, and can automatically adjust the size and step size of the pooling kernel according to the importance of the feature, adapting to feature representations of different scales and resolutions; inputs the fused feature representation into the deep metric network for classification and recognition, and can learn the similarity or distance between different categories, thereby achieving more accurate classification and recognition of radar signal patterns. Combining the above advantages, the method of the present invention can make full use of the information of multimodal features, improve the quality of feature extraction and the accuracy of classification and recognition, has strong adaptability and robustness, and is suitable for radar signal recognition tasks in complex and changing environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 is a flow chart of the method of the present invention;

[0032] Figure 2 It is a principle block diagram of the 1DResNet network in the present invention;

[0033] Figure 3 It is a principle block diagram of the 2DResNet network in the present invention;

[0034] Figure 4 It is a schematic diagram of the adaptive feature fusion method in the present invention;

[0035] Figure 5 It is a schematic diagram for comparing the performance of the methods provided in the embodiments of the present invention. DETAILED DESCRIPTION

[0036] The present invention is further explained below in conjunction with the accompanying drawings and specific embodiments. It should be understood that these embodiments are only used to illustrate the present invention and are not used to limit the scope of the present invention. After reading the present invention, various equivalent forms of modifications to the present invention by those skilled in the art all fall within the scope defined by the claims attached to this application.

[0037] The present invention provides a radar signal pattern recognition method based on multi-modal feature fusion, such as Figure 1 As shown, it includes the following steps:

[0038] S1: The IQ sampling data and full pulse PDW data of the radar signal are acquired through the reconnaissance receiver, and the IQ sampling data is transformed into a time-frequency image; the time-frequency transformation methods include short-time Fourier transform, wavelet transform, Wigner-Ville distribution and pseudo-Wigner-Ville distribution, etc. The transformation method with higher time-frequency resolution and better suppression of cross terms is selected to obtain the time-frequency image of the signal;

[0039] S2: preprocessing the time-frequency image and full-pulse PDW data respectively;

[0040] The preprocessing method of the time-frequency image is: The preprocessing method of the time-frequency image is: Because the time-frequency image is a color image, converting the color image to a grayscale image can simplify the data processing process, reduce the demand for computing resources, improve the robustness of the algorithm, and in many cases help to highlight the key information in the image more clearly. Therefore, based on the brightness level of the image pixel, a weighted sum relationship between the brightness (Luminance) Y and the three color channels of red (R), green (G), and blue (B) is established.

[0041] Specifically, based on the brightness level of the image pixels, the corresponding relationship between brightness Y and the three color channels of R, G, and B is established: Y = 0.2989R + 0.5870G + 0.1140B, the time-frequency image is denoised, the noise of the grayscale image is filtered out using the smoothing filter method, and the denoised image is converted into a grayscale image using the weighted average method. The core of this method is to eliminate the hue and saturation information of the image, express the grayscale value of the image as a brightness value, and realize the grayscale processing of the image.

[0042] The preprocessing method of the full-pulse PDW data is: expanding the full-pulse PDW data of the radar signal, that is, expanding the PDW data into a one-dimensional feature vector.

[0043] The present invention adopts a method based on the sampling frequency f s The uniform distribution U(·) is used to uniformly represent the parameter value of the signal. For example, U(1 / 5,1 / 3) represents the parameter range [f s / 5,f s / 3] within the random number.

[0044] S3: extract features from the preprocessed time-frequency image through a high sampling rate intra-pulse IQ data feature extraction network to obtain intra-pulse features; extract features from the preprocessed full-pulse PDW data through a low sampling rate PDW data feature extraction network to obtain inter-pulse features;

[0045] like Figure 2 As shown in the figure, the low sampling rate PDW data feature extraction network is a 1DResNet network. The network extracts features through 1 1D convolution layer, 3 residual modules and 1 fully connected layer. The 3 residual modules have the same architecture. Each residual module contains 2 1D convolution layers and corresponding batch normalization layers and activation functions. Residual connections are added between modules to adapt to deeper networks; each convolution layer is equipped with 1 batch normalization layer and 1 ReLU activation function for adaptively adjusting the changing gradient; after the last residual module, a global average pooling is used to process input data of different lengths so that the output features are consistent, to extract important inter-pulse features of each channel, and obtain the image feature matrix as the inter-pulse features.

[0046] like Figure 3 As shown in the figure, the high sampling rate intra-pulse IQ data feature extraction network is a 2DResNet network. The network extracts features through 1 2D convolution layer and 3 residual modules. The 3 residual modules have the same architecture. Each residual module contains 2 2D convolution layers and corresponding batch normalization layers and activation functions. Residual connections are added between modules to adapt to deeper networks. Each convolution layer is equipped with 1 batch normalization layer and 1 ReLU activation function for adaptively adjusting the changing gradient. After the last residual module, a global average pooling is used to process input data of different lengths so that the output features are consistent, to extract important inter-pulse features of each channel, and obtain the image feature matrix as the intra-pulse features.

[0047] It should be noted here that the one-dimensional convolution layer (Conv1d) is usually used to process one-dimensional data. A parameter sequence of PDW data is a one-dimensional sequence. These data have a main feature dimension. The two-dimensional convolution layer (Conv2d) is usually used to process data with a two-dimensional spatial structure, such as a transformed time-frequency grayscale image. The two-dimensional data has height and width (such as rows and columns of an image). Because the input features are different, the convolution layer used will be different.

[0048] S4: Adaptive fusion of multimodal features for intra-pulse features and inter-pulse features;

[0049] The adaptive feature fusion method is:

[0050] A1: Adaptive pooling operations are applied to inter-pulse features and intra-pulse features respectively to obtain the pooling results of inter-pulse features and intra-pulse features;

[0051] Adaptive pooling: A global average pooling is used after the last residual module, which can handle input data of different lengths and make the output features consistent.

[0052] A2: Fuse the pooling results of inter-pulse features and intra-pulse features to obtain a fused feature vector;

[0053] like Figure 4 As shown, two learnable parameters α and β are defined in the present invention to weight the feature vectors of 1D ResNet and 2D ResNet, and the fused feature vector is expressed as:

[0054] Fused Feature=α×1D? Feature+β×2D? Feature (1)

[0055] To ensure that the weights sum to 1, the Softmax function is used to normalize the weights:

[0056] α,β=softmax([w1,w2]) (2)

[0057] Where w1 and w2 are two learnable parameters.

[0058] In the present invention, the weighted fused feature vector is calculated and classified through a fully connected layer, and the model parameters are optimized by back propagation using a standard training procedure; the weights of the 1D ResNet and 2D ResNet feature vectors are dynamically adjusted by formula (1), so that the model can automatically learn the best feature fusion method according to the data.

[0059] S5: Use the trained deep metric network as a classifier, input the fused feature vector into the deep metric network, learn the feature similarity metric space, output the final recognition result, and realize radar signal style recognition.

[0060] Based on the above content, in order to verify the effectiveness of the method of the present invention, this embodiment has been simulated and verified, as follows:

[0061] 1. Simulation conditions:

[0062] The simulation was carried out on an Intel(R) Core(TM) i5-8250U CPU@1.60GHz and WINDOWS10 system. MATLAB was used for data simulation, and the experiment was run on PyCharm 2022.

[0063] 2. Simulation content:

[0064] Since it is difficult to obtain actual radar signals, experimental radar data is generated through simulation, as shown in Table 1, mainly including RF, PW, PRI features and different intra-pulse modulation types. Among them, NLFM modulation uses sinusoidal modulation, and BPSK modulation uses 7-bit Barker code. The simulation sets each function of a radar to correspond to a specific radar signal style of a combination of intra-pulse and inter-pulse features. PDWS mainly describes the radar working mode, including speed search (VS), target tracking (STT), search mode (STT) and search plus tracking (TAS). Each signal sample is randomly generated within a given range, not limited to a specific model. Each working mode sample is randomly combined from a given parameter range. Four typical radar intra-pulse modulation types are considered, with a step size of 2dB in the SNR range of -8dB to 6dB, and 80 time-frequency images are generated for each step size. According to the function expressed by the radar, the radar intra-pulse modulation type and working mode are randomly combined, and four types of radar signal style combinations are assigned to each type of radar.

[0065] Table 1 Radar signal style and parameter settings

[0066]

[0067]

[0068] In order to verify the effectiveness of the neural network in the learning framework, the training was performed with 80% of the training set and 20% of the validation set, and the test set was completely independent of the training set and the validation set. In total, 960 time-frequency images were generated for each SNR.

[0069] As shown in Table 1, the three radars have a total of 12 radar signal patterns, and twelve labels are assigned to them. Three experimental groups with different numbers of samples are set up for them. 90%, 50%, and 10% of the samples are taken from the training set for training at each signal-to-noise ratio, corresponding to sample one, which corresponds to 72 time-frequency images of each radar signal pattern; sample two, which corresponds to 40 time-frequency images of each radar signal pattern; sample three, which corresponds to 8 time-frequency images of each radar signal pattern. The radar signal pattern data is subjected to comparative experiments with different numbers of samples using the method of the present invention. By constructing a confusion matrix and calculating the number of correctly predicted samples, the overall accuracy is obtained, and the performance of radar signal pattern recognition accuracy at different signal-to-noise ratios is compared. The final comparison is shown as follows. Figure 5 shown.

[0070] from Figure 5It can be seen that at low signal-to-noise ratios above -6dB, the three groups of experiments with different sample numbers have an accuracy rate greater than 70%; when the signal-to-noise ratio is greater than 2dB, the three groups of experiments can obtain an accuracy rate close to 100%, indicating that the recognition accuracy of the present invention is quite impressive. Sample three has the least number, but after the signal-to-noise ratio is greater than 0dB, the recognition rate is close to that of sample two and sample one, and the recognition accuracy of sample two and sample one is always close. This shows that the present invention still has good performance in small sample signal pattern recognition.

Claims

1. A radar signal pattern recognition method based on multimodal feature fusion, characterized in that: The steps include: S1: Obtain the IQ sampling data and full pulse PDW data of the radar signal pulse, and perform time-frequency transformation on the IQ sampling data to obtain a time-frequency image; S2: preprocessing the time-frequency image and full-pulse PDW data respectively; S3: extracting features from the preprocessed time-frequency image and full-pulse PDW data respectively to obtain intra-pulse features and inter-pulse features; S4: Adaptive fusion of multimodal features for intra-pulse features and inter-pulse features; S5: Use the fused features as classifier input data for classification and recognition to achieve radar signal pattern recognition.

2. The radar signal pattern recognition method based on multimodal feature fusion according to claim 1 is characterized in that: The preprocessing method of the time-frequency image in step S2 is: based on the brightness level of the image pixel, a corresponding relationship between the brightness Y and the three color channels of R, G, and B is established: Y=0.2989R+0.5870G+0.1140B, the time-frequency image is denoised, the noise of the grayscale image is filtered out using a smoothing filtering method, and the denoised image is converted into a grayscale image using a weighted average method.

3. The radar signal pattern recognition method based on multimodal feature fusion according to claim 1 is characterized in that: The preprocessing method of the full-pulse PDW data in step S2 is: expanding the full-pulse PDW data of the radar signal, that is, expanding the PDW data into a one-dimensional feature vector.

4. The radar signal pattern recognition method based on multimodal feature fusion according to claim 1 is characterized in that: In step S3, feature extraction is performed on the preprocessed time-frequency image through a high sampling rate intra-pulse IQ data feature extraction network to obtain intra-pulse features; and feature extraction is performed on the preprocessed full-pulse PDW data through a low sampling rate PDW data feature extraction network to obtain inter-pulse features.

5. The radar signal pattern recognition method based on multimodal feature fusion according to claim 4 is characterized in that: The high sampling rate intra-pulse IQ data feature extraction network is a 2DResNet network. The network extracts features through 1 2D convolution layer and 3 residual modules. The 3 residual modules have the same architecture. Each residual module contains 2 2D convolution layers and corresponding batch normalization layers and activation functions. Residual connections are added between modules to adapt to deeper networks; each convolution layer is equipped with 1 batch normalization layer and 1 ReLU activation function for adaptively adjusting the changing gradient; after the last residual module, a global average pooling is used to extract important inter-pulse features of each channel to obtain an image feature matrix as the intra-pulse feature.

6. The radar signal pattern recognition method based on multimodal feature fusion according to claim 4 is characterized in that: The low sampling rate PDW data feature extraction network is a 1DResNet network, which extracts features through 1 1D convolution layer, 3 residual modules and 1 fully connected layer. The 3 residual modules have the same architecture. Each residual module contains 2 1D convolution layers and corresponding batch normalization layers and activation functions. Residual connections are added between modules to adapt to deeper networks; each convolution layer is equipped with 1 batch normalization layer and 1 ReLU activation function for adaptively adjusting the gradient of changes; after the last residual module, a global average pooling is used to extract important inter-pulse features of each channel, and an image feature matrix is ​​obtained as the inter-pulse features.

7. The radar signal pattern recognition method based on multimodal feature fusion according to claim 1 is characterized in that: The adaptive fusion method in step S4 is: A1: Adaptive pooling operations are applied to inter-pulse features and intra-pulse features respectively to obtain the pooling results of inter-pulse features and intra-pulse features; A2: Fuse the pooling results of inter-pulse features and intra-pulse features to obtain a fused feature vector.

8. The radar signal pattern recognition method based on multimodal feature fusion according to claim 7 is characterized in that: The step A2 is specifically as follows: Define two learnable parameters α and β to weight the feature vectors of 1D ResNet and 2D ResNet. The fused feature vector is expressed as: Fused Feature=α×1D? Feature+β×2D? Feature (1) To ensure that the weights sum to 1, the Softmax function is used to normalize the weights: α,β=softmax([w1,w2]) (2) Where w1 and w2 are two learnable parameters.

9. The radar signal pattern recognition method based on multimodal feature fusion according to claim 1, characterized in that: In step S5, the classifier is a trained deep metric network, and the fused feature vector is input into the deep metric network to learn the feature similarity metric space and output the final recognition result.

Images