PRI type classification method for driving continuous small-band waves based on attention mechanism

Abstract

The invention belongs to the technical field of signal processing, and particularly relates to a PRI type classification method for driving continuous small-band waves based on an attention mechanism, and the method comprises the following steps: CWT is used for decomposing a PRI modulation sequence and obtaining different time-frequency components, and features extracted from a CWT 2D scale map are used for executing PRI modulation discrimination by means of a proposed CNN-MHSA-LSTM combination. Wherein the vanilla CNN extracts depth features to find class details, and captures spatial attributes at the same time. In order to improve the discrimination capability of the whole framework, an MHSA mechanism is used. The time attribute is obtained by the LSTM, and the LSTM works in cooperation with the CNN, and carries out the detection of the PRI category according to the extracted features. A simulation result shows that when the environment fluctuates and the stray pulse is 50%, the optimal identification accuracy is 95.33%.

Description

Technical Field

[0001] The present invention belongs to the field of passive detection and involves technologies such as radar signal recognition and artificial intelligence deep learning algorithms. Background Art

[0002] The electromagnetic (EM) environment has become more crowded due to the use of complex communication, navigation, and radar systems. Electronic warfare (EW) continues to evolve and become more complex, attempting to control and dominate the radio frequency (RF) spectrum. Pulse repetition interval modulation identification is an important part of electronic intelligence signal analysis, and its accuracy plays an important role in improving radar performance. PRI (Pulse Repetition Interval,) modulation types are generally classified as fixed, staggered, jittered, sinusoidal modulation, dwell switching, and sliding. Pulse repetition interval (PRI) modulation type radar signals are inherently characteristic of radar emitters and their operating modes. PRI describes the gap pattern between the leading edges of subsequent radar pulses. Radar PRI modulation patterns are an important part of ES system emitter identification and can be used to infer the purpose and application of the radar emitter. Therefore, understanding the PRI modulation type is essential to identifying radar emitters and their operating modes. However, with the improvement of countermeasures and the proliferation of radar capabilities, PRI modulation has become more challenging. The complexity is further increased when PRI is intertwined with background noise. Therefore, some traditional methods, such as statistical-based approaches, decision tree-based techniques, and learning-based procedures using shallow artificial neural networks (ANNs), may not be sufficient to accurately identify various PRI modulation types in noisy environments. Summary of the invention

[0003] In order to solve the problem of unsatisfactory PRI accuracy in the prior art, the present invention provides a PRI type classification method based on attention mechanism driving continuous small waves.

[0004] The present invention proposes a PRI type classification method based on attention mechanism driven continuous small wave, and the technical solution includes:

[0005] Step 1: Use MATLAB or Python to simulate and generate eight types of complex PRI modulation sequences, including sequence distortion effects such as Gaussian white noise, speckle noise, impulse noise, burst noise, and noise from imperfect channels.

[0006] Step 2: Perform data feature engineering based on the simulated PRI data: extract dimensional data, process missing values ​​for the dimensional data, and remove redundant data; divide the simulated data set into a training set, a cross-validation set, and a test set in a ratio of 7:2:1. The training set helps train the model parameters, the cross-validation set helps determine the excellence of the parameters, and finally the test set is used to test the accuracy of the model.

[0007] Step 3: Use continuous wavelet transform (CWT) to extract time-frequency information during training. A combination model of convolutional neural network (CNN), multi-head self-attention mechanism (MHSA) and popular long short-term memory (LSTM) is used to improve the accuracy of automatic recognition of several complex PRI modulation types. The present invention uses continuous wavelet transform to obtain the time-frequency properties of PRI sequences in the form of two-dimensional scalograms, and uses these time attribute features to train the model. A CNN-LSTM model based on end-to-end hybrid MHSA is built to capture and enhance the spectral and temporal properties of the applied samples, thereby improving recognition performance.

[0008] Step 4: Facilitate full end-to-end learning for PRI modulation type discrimination in a random noisy electromagnetic environment background.

[0009] Step 5: Use the cross-validation set to adjust the model parameters of the trained model; find the optimal model and test the model recognition results through the final test set.

[0010] For benchmark comparison, multiple models were implemented and trained in a similar way. This situation can be overcome by deep learning techniques such as CNN that have high tolerance to noise-induced changes. To evaluate the effectiveness of the proposed method, three models based on ResNet, popular CNN, and SqueezeNet were implemented to benchmark in terms of overall performance and complexity. Finally, the recognition accuracy of the proposed method is better than the above three models.

[0011] The beneficial effects of the present invention include:

[0012] The accuracy of the PRI modulation type classification based on the present invention is better than that of popular models such as ResNet, popular CNN and SqueezeNet, and the parameters of the model are in the same data set as other models. The present model increases the recognition accuracy of PRI while maintaining the original performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] Figure 1 Main processing flow chart of classification PRI recognition algorithm software.

[0014] Figure 2 A schematic diagram of PRI types according to an embodiment of the present invention. DETAILED DESCRIPTION

[0015] The technical solutions in the embodiments of the present invention will be further described in detail below in conjunction with the accompanying drawings in the embodiments of the present invention.

[0016] The implementation process of the present invention is as follows Figure 2As shown, first, eight types of complex PRI modulation sequences are generated by matlab or python simulation, and the sequence distortion effects include Gaussian white noise, speckle noise, impulse noise, burst noise and noise of imperfect channels. The present invention takes into account eight types of PRI modulation used in radar systems in the prior art: Generally, sliding and ripple PRI are called complex PRI modulation types, and subtypes of these complex PRI modulation types include sawtooth wave, sine wave, saturated sine wave and triangle wave. The curves of different PRI modulation types under ideal conditions are shown in Figure 1 As shown. In a real electronic warfare environment, PRI is susceptible to illusions and spurious pulses, resulting in considerable distortion of the original PRI sequence form, making the classification task extremely difficult. With the advancement of radar technology, different radar systems use a variety of complex PRI modulation modes. The PRI sequence of a radar pulse signal is defined as: P(j) = (j+1)-t(j), j = 0, 1, 2, ...., L-1 (1) where 'L' is the number of pulses detected and 't(j)' represents the estimated arrival time of the jth radar pulse. The mode of the PRI sequence is controlled by a specific PRI modulation type, and different PRI modulations have different functions in radar systems.

[0017] The next step is to perform data feature engineering on the PRI data simulated by the simulation program: First, simulate the experimental data through the code, and land the simulated raw data on the server. Perform data cleaning on the collected raw data: Use visualization methods to display, analyze, process, and clean the parameters of the original detection data. Under this condition, perform missing value processing on the collected dirty data, remove redundant data, and repair abnormal data, extreme value data, etc. Standardize and normalize the cleaned data: During the training of the neural network, due to the scattered distribution of sample features, the deep neural network may learn slowly, and even cause problems such as gradient vanishing and gradient explosion in severe cases. Use the batch normmailzation method to normalize the input samples with radar features so that the data has a mean of 0 and a variance of 1. In other words, it obeys a normal distribution. The preferred algorithm is as follows:

[0018]

[0019] After processing by the above algorithm, we get training data that obeys normal distribution. The normal distribution is as follows:

[0020]

[0021] According to the traditional deep learning training method, the present invention divides the simulation data set into a training set, a cross-validation set and a test set in a ratio of 7:2:1. The training set helps to train the model parameters, the cross-validation set helps to determine the excellence of the parameters, and finally the test set is used to test the accuracy of the model.

[0022] Next, a method for detecting PRI modulation using MHSA-based CNN-LSTM and CWT is used. The first step of the method is to generate a 2D scale map using CWT with a series of changes in the PRI sequence as described above, and provide the processed scale map as input to the deep learning structure for classification. The neural network is mainly used to extract relevant features for PRI sequence samples, including: In the architecture, CNN is used to extract deep features restricted to spatial attributes, and LSTM is used to fine-tune time segments to use these relevant contents for classification. As emphasized before, in order to improve the effectiveness of the model, the MHSA mechanism is combined. First, the PRI sequence is converted into its time-frequency representation using CWT. These scale maps are then used as input to the CNN. In order to provide a clearer representation, the convolution, batch normalization, and ReLU layers are merged into a convolutional layer unit. The network has a total of 24 layers: four convolutional layers, three maximum pooling layers, one FC layer, one MHSA layer, one LSTM layer, and an output layer with a softmax activation function. Features are extracted from the scale map using four consecutive convolution and maximum pooling operations. Along the convolutional layer, the number of filters increases from 8 to 64, and the filter size is constant at 3×3. The weights of the convolution kernels are trained using the back propagation (BP) error algorithm. ReLU is used as the nonlinear activation function. The PRI sequence is periodic, i.e., PRI values ​​appear in a repetitive manner. Therefore, the PRI values ​​contained in a given period can be used to determine its type. This inference is based on intuition. In addition, due to the presence of spurious pulses in the PRI sequence, a portion of PRI values ​​belongs exclusively to the initial PRI sequence and is suitable for recognition purposes. Incorrect PRI values ​​do not have any impact on the calculation of class assignment.

[0023] Secondly, incorporating the attention mechanism into the CNN module enables the model to selectively focus on salient features during training, thereby improving the training efficiency and the classification accuracy of the model. The MHSA mechanism is used to assign weights to the features extracted by the previous network, so that important features can be distinguished. Finally, the design and operation of the benchmark model are implemented to compare with the proposed CNN-MHSA-LSTM method. A CNN neural network is implemented, which has a total of 27 layers: six convolutional layers with ReLU activation function, five maximum pooling layers, one FC, and an output layer with softmax activation function. In addition, a batch normalization (BN) layer is used between each convolutional layer and ReLU layer to reduce training time. The CWT dataset is used for training and testing. The experimental results show that the commonly used CNN has some room for improvement in performance when processing CWT samples of the above PRI modulation types.

[0024] For benchmark comparison, multiple models are implemented and trained following a similar approach. This situation can be overcome by deep learning techniques such as CNN that have a high tolerance to noise-induced changes. To evaluate the effectiveness of the proposed method, three models based on ResNet, popular CNN, and SqueezeNet were implemented. Residual CNN Network: DNN with residual connections or residual networks (ResNet) is a model developed to overcome the confusion in deep learning training. SqueezeNet uses certain design strategies to reduce the number of parameters while maintaining accuracy. To perform a benchmark comparison in terms of overall performance and complexity. In the end, the recognition accuracy of the present invention is better than the above three models.

Claims

1. PRI type classification method based on attention mechanism driven continuous small wave: It is characterized by: Step 1: Generate eight types of complex PRI modulation sequences using matlab or python simulation, and their sequence distortion effects; wherein the sequence distortion effects include Gaussian white noise, speckle noise, impulse noise, burst noise and noise of imperfect channels; Step 2: Perform data feature engineering based on the simulated PRI data: extract dimensional data, process missing values ​​for the dimensional data, and remove redundant data; divide the simulated data set into a training set, a cross-validation set, and a test set in a ratio of 7:2:1; use the training set to train model parameters; Step 3: During the training process, continuous wavelet transform is used to obtain the time-frequency properties of the PRI sequence in the form of a two-dimensional scalogram; Build a CNN-LSTM model based on a hybrid multi-head self-attention mechanism to capture and enhance the spectral and temporal properties of the applied samples; Step 4: Facilitate full end-to-end learning for PRI modulation type discrimination in a random noise electromagnetic environment background; Step 5: Use the cross-validation set to adjust the model parameters of the trained model; find the optimal model and test the model recognition results through the final test set.

Images