Signal recognition method based on time-frequency analysis and deep learning under DRFM intermittent sampling

Abstract

The disclosed signal recognition method based on time-frequency analysis and deep learning under DRFM intermittent sampling belongs to the field of radio fuze countermeasure. The implementation method of the present application is as follows: a radio fuze jammer based on digital radio frequency memory (DRFM) uses a time-sharing antenna to obtain target signals and retransmit interference signals; the radio fuze signal is obtained by using the intermittent sampling mode; the time-frequency image of the target signal is extracted by using the smooth pseudo-Wigner-Ville distribution (SPWVD), the generation of cross terms is significantly suppressed, the signal preprocessing effect of automatic modulation recognition is improved, and the distinguishability of the signal is improved; a radio fuze signal automatic modulation recognition model based on a Vision Transformer backbone network is constructed; the automatic modulation recognition model is trained as a classifier by using the pre-training and fine-tuning mode, the received radio fuze signal is sorted and recognized, and then the sorting and recognition of various radio fuze signals can be realized under the intermittent sampling scene such as DRFM interference.

Description

Technical Field

[0001] This invention relates to a signal recognition method based on time-frequency analysis and deep learning under DRFM intermittent sampling, belonging to the field of radio fuze countermeasures. Background Technology

[0002] With the development of signal transmission technology, signals of different frequencies and modulation types permeate the electromagnetic space, forming a complex battlefield electromagnetic environment. Radio fuses are special devices that utilize the electromagnetic scattering characteristics of targets to detect them and control the detonation of munitions. Electronic countermeasures against radio fuses are a dynamic game, with both sides engaging in fierce electromagnetic spectrum competition within the complex electromagnetic space. Currently, DRFM jamming is the mainstream jamming method in electronic countermeasures against radio fuses. With the widespread adoption of DRFM jamming technology, many scholars have proposed various anti-jamming methods to eliminate the effects of jamming and protect the integrity of target echo information. For example, alternating the modulation rate of the frequency-modulated fuse's transmitted signal can improve the DRFM fuse's resistance to jamming. However, with the development of radio fuse anti-jamming technology, the jamming effect of current single-strategy DRFM jamming techniques will inevitably decrease.

[0003] Introducing cognitive radio technology into the jamming process can improve the cognitive level and environmental awareness of radio fuse jammers, and provide information support for jamming decision-making. Automatic modulation identification (AMI) is a commonly used cognitive radio technology that can identify the modulation type of blind source signals. In the jamming process of radio fuses, using AMII to obtain the modulation type information of the target signal and dynamically adjusting the jamming strategy and method according to the signal modulation type can improve the cognitive jamming level and jamming effect. AMII is widely used in spectrum sensing, blind source signal classification, and jamming protection. In 2018, S. Liu (“Radar Emitter Recognition Based on SIFT Position and Scale Features”) proposed a method based on Scale Invariant Feature Transform (SIFT) to extract the modulation features of signals, and used support vector machines to classify the extracted features, achieving modulation type identification for multiple signals. However, AMII methods that manually extract signal features rely heavily on prior knowledge. When the number of signal modulation types increases and the modulation patterns become more complex, they cannot achieve good recognition results. In 2022, H. Yu (“Radar emitter multi-label recognition based on residual network”) extracted the normalized time-frequency image of the signal using short-time Fourier transform, then applied a deep denoising model to denoise the image, and finally input the denoised time-frequency image into a residual neural network for classification, achieving signal modulation type recognition at a low signal-to-noise ratio. However, current automatic modulation recognition methods are not suitable for applications involving intermittent sampling during DRFM jamming. In electronic countermeasures using radio fuses with DRFM jamming, the receiver acquires the target signal through intermittent sampling. The target signal exhibits time-frequency domain truncation characteristics. Compared to signals used for automatic modulation recognition in general scenarios, signals obtained through intermittent sampling in DRFM jamming scenarios are more complex, carry less information, and are more difficult to identify in terms of modulation type. Summary of the Invention

[0004] This invention addresses the problem of signal modulation type identification in intermittent sampling scenarios such as DRFM interference in radio fuses. The main objective is to provide a signal identification method based on time-frequency analysis and deep learning under DRFM intermittent sampling. This method utilizes intermittent sampling to acquire radio fuse signals; extracts the time-frequency image of the target signal using a smoothed pseudo-Wigner-Villi distribution (SPWVD); constructs an automatic modulation identification model for radio fuse signals based on a Vision Transformer backbone network; and trains the automatic modulation identification model as a classifier through pre-training and fine-tuning to sort and identify the received radio fuse signals. This enables the sorting and identification of various radio fuse signals even in intermittent sampling scenarios such as DRFM interference.

[0005] The objective of this invention is achieved through the following technical solution.

[0006] The signal recognition method based on time-frequency analysis and deep learning under DRFM intermittent sampling disclosed in this invention includes the following steps:

[0007] Step 1: The radio fuse jammer based on Digital Radio Frequency Storage (DRFM) acquires the time-frequency truncated radio fuse signal using an intermittent sampling mode according to different jamming and forwarding strategies, and obtains the down-converted target signal through down-conversion processing.

[0008] The radio fuse jammer based on Digital Radio Frequency Storage (DRFM) employs a time-division multiplexing antenna to acquire target signals and relay jamming signals. During the transmission of jamming signals, no signals are received. Therefore, the DRFM jamming equipment operates in an intermittent sampling and jamming relay mode.

[0009] Target signals obtained through intermittent sampling exhibit time-frequency domain truncation. The sampling method of intermittent sampling is related to the jamming forwarding strategy. There are three DRFM jamming forwarding strategies: direct forwarding, repeated forwarding, and cyclic forwarding. Different jamming forwarding strategies result in different time-domain signal truncation effects and signal lengths. The signal truncation patterns under different jamming forwarding strategies have the following characteristics: In the intermittent sampling direct forwarding process, after receiving the target signal, the radio fuse jammer directly forwards the target signal. At this time, the time-domain signal obtained through intermittent sampling is discontinuous and continuous. In the intermittent sampling repeated forwarding process, after receiving the target signal, the radio fuse jammer first repeatedly splices the same signal segment multiple times, and then forwards the spliced ​​jamming signal. In the intermittent sampling cyclic forwarding process, after receiving the target signal, the radio fuse jammer sequentially splices the previously obtained signals, and then forwards the spliced ​​jamming signal. In the cyclic forwarding process, the signal splicing process is repeated periodically, and the signal stored in the previous cycle is not forwarded.

[0010] Depending on the different interference forwarding strategies, the radio fuze receiver acquires the time-frequency truncated target signal through intermittent sampling, and performs down-conversion processing on the target signal to obtain the down-converted target signal.

[0011] Step 2: Using the smoothed pseudo-Wigner-Villi distribution SPWVD, the target signal after frequency conversion processing in Step 1 is windowed and smoothed in both the time and frequency domains to suppress the generation of cross terms, enhance the time-frequency characteristics of the signal, and generate the SPWVD time-frequency image of the signal.

[0012] SPWVD performs time-domain and frequency-domain windowing smoothing on the signal, suppressing the generation of cross terms and enhancing the time-frequency characteristics of the signal.

[0013] For a discrete signal x(n), where n = 0, 1, 2, ..., N-1, its SPWVD operation expression is:

[0014]

[0015] Where t(m) represents the smoothing window function in the time domain and g(k) represents the smoothing window function in the frequency domain, both of which are Hamming windows.

[0016] After obtaining the time series of the radio fuze signal by intermittent sampling, the time-frequency analysis method of Equation (1) is used to perform time-frequency analysis on the time series to obtain the SPWVD time-frequency image of the signal.

[0017] Step 3: Establish an automatic modulation and recognition model for radio fuze signals based on the Vision Transformer backbone network. This model consists of an encoder, a decoder, and an output layer, with the encoder and decoder using the Masked Autoencoders algorithm structure. The time-frequency image is input into the encoder, which performs positional encoding and random masking on the image, extracts features from the time-frequency image using the Transformer structure, and outputs a feature map. The decoder reconstructs the time-frequency image by fitting the difference between the feature map and the original time-frequency image. The feature extraction and image reconstruction capabilities of the Vision Transformer-based automatic modulation and recognition model for radio fuze signals are trained by minimizing the difference between the reconstructed and original time-frequency images. After reconstructing the time-frequency image, the decoder outputs a one-dimensional feature vector to the output layer. The output layer performs dimensionality reduction and classification on the one-dimensional feature vector to obtain a one-dimensional recognition vector representing the modulation type of the radio fuze signal. The training process of the Vision Transformer-based automatic modulation and recognition model for radio fuze signals adopts a pre-training plus fine-tuning training method to reduce the training cost and improve the training efficiency. During pre-training, time-frequency reconstruction training is performed using complete, unlabeled time-frequency images of radio fuze signals, without modulation type identification. The fine-tuning training process first transfers the parameters of the pre-trained Vision Transformer-based automatic modulation identification model of radio fuze signals to the fine-tuning model via parameter transfer, initializing the fine-tuning model. Then, a small number of time-frequency images of radio fuze signals acquired through intermittent sampling are used to train the Vision Transformer-based automatic modulation identification model for modulation type identification.

[0018] Step 3.1: Establish an automatic modulation and identification model for radio fuze signals based on the Vision Transformer backbone network;

[0019] The automatic modulation and identification model for radio fuze signals based on Vision Transformer consists of an encoder, a decoder, and an output layer. The encoder and decoder use the Masked Autoencoders algorithm structure and are both cascaded Transformer blocks. The encoder and decoder are connected sequentially.

[0020] To balance model complexity with recognition performance, reduce computational cost, and improve recognition accuracy, the encoder preferably consists of 12 layers of Transformer blocks, with 12 multi-head attention heads. The decoder consists of 8 layers of Transformer blocks, with 16 multi-head attention heads.

[0021] The Transformer block is entirely based on attention mechanisms, without using any convolutional layers or recurrent neural network layers. A Transformer block contains two normalized Linear Neural Network (LN) layers, a multi-head attention layer, and a multilayer perceptron. The first LN layer is cascaded with the multi-head attention layer to form a residual structure, and the second LN layer is cascaded with the multilayer perceptron to form another residual structure; these two residual structures are sequentially connected. Within the Transformer block, the multi-head attention layer consists of multiple self-attention layers and a fully connected layer. The fully connected layer aggregates the outputs of each self-attention layer to obtain the overall output. Multi-head attention enhances the network's stability and robustness. The multilayer perceptron consists of two fully connected layers, a Gaussian error linear unit activation function (GELU) layer, and two dropout layers. By randomly discarding some features through the dropout layers, the network's robustness is improved, suppressing overfitting. GELU is a variant of the ReLU linear rectified function. The GELU activation function has better non-linear performance than the ReLU activation function, significantly improving model performance.

[0022] The GELU activation function is expressed as follows:

[0023]

[0024] The output layer, connected to the decoder, consists of a fully connected layer and a Softmax activation function layer. The decoder's output, after passing through the fully connected layer, yields a one-dimensional probability vector of length N, where N represents the total number of modulation types for the radio fuze signal. The Softmax activation function layer then processes this one-dimensional probability vector to maximize the output probability. The Softmax activation function transforms unnormalized predictions into non-negative numbers that sum to 1. The Vision Transformer-based automatic modulation recognition model for radio fuze signals outputs a one-dimensional recognition vector composed of 0s and 1s, representing the modulation type of the radio fuze signal.

[0025] Step 3.2: Train the automatic modulation and recognition model of radio fuze signals based on Vision Transformer using pre-training and fine-tuning to reduce the training cost of the model and improve its recognition performance.

[0026] The Vision Transformer-based automatic modulation recognition model for radio fuze signals was trained using a pre-training and fine-tuning approach. The pre-training data used were unlabeled, continuous, and complete time-frequency images. The automatic modulation recognition model simulated intermittent sampling by randomly discarding portions of the time-frequency image. During pre-training, the loss function used was the mean squared error (MSE).

[0027]

[0028] Where Y i Represents the actual sample distribution. This represents the predicted sample distribution.

[0029] The model's feature extraction capability is trained by fitting the difference between the time-frequency image generated by the decoder and the actual time-frequency image through time-frequency reconstruction. Since the pre-training process uses unlabeled, complete time-frequency images, modulation type identification of the radio fuze signal is not performed during pre-training.

[0030] During fine-tuning training, the initial parameters of the model are transferred from the pre-trained model. The training data used for fine-tuning is a small amount of labeled time-frequency images obtained through intermittent sampling. The loss function used in fine-tuning training is the cross-entropy loss function.

[0031]

[0032] Where y i Represents the actual category. This represents the category of the prediction.

[0033] Cross-entropy loss is used in multi-class classification tasks to calculate the difference between predicted and true values. The model's ability to identify radio fuze modulation types is trained by fitting the error between the predicted modulation type and the true modulation type of the radio fuze signal predicted by the Vision Transformer-based automatic modulation identification model. The optimization algorithm used in the pre-training and fine-tuning processes is AdamW. AdamW improves the generalization performance of the Vision Transformer-based automatic modulation identification model by increasing regularization weights while maintaining the model's fast convergence ability.

[0034] Step 4: Input the SPWVD time-frequency diagram from Step 2 into the VisionTransformer-based automatic modulation and identification model for radio fuze signals trained in Step 3. The model sorts the signal modulation type and identifies radio fuze signals of different systems, thereby improving the accuracy and efficiency of radio fuze signal identification in DRFM intermittent sampling scenarios.

[0035] The specific steps for signal sorting and identification of the SPWVD time-frequency diagram in step two using the Vision Transformer-based automatic modulation and identification model for radio fuze signals are as follows:

[0036] Step 4.1: The encoder of the automatic modulation and recognition model for radio fuze signals based on Vision Transformer segments the time-frequency image of the input radio fuze signal into image blocks of equal size;

[0037] Step 4.2: The encoder performs position encoding on all image blocks and randomly removes some image blocks;

[0038] Step 4.3: The encoder extracts features from the remaining image patches using the Transformer structure;

[0039] Step 4.4: Pad the missing parts of the feature map with zeros according to the position encoding;

[0040] Step 4.5: The decoder of the automatic modulation and identification model of the radio fuze signal reconstructs the original time-frequency image based on the zero-padded feature map and outputs a one-dimensional feature vector;

[0041] Step 4.6: The output layer of the automatic modulation and identification model for radio fuze signals performs dimensionality reduction and classification on the one-dimensional feature vector output by the decoder, and maximizes the prediction probability.

[0042] Step 4.7: The output layer outputs a one-dimensional identification vector representing the modulation type of the radio fuze signal, realizing the sorting and identification of radio fuze signals of different systems.

[0043] Beneficial effects:

[0044] 1. The present invention discloses a signal recognition method based on time-frequency analysis and deep learning under DRFM intermittent sampling. In order to improve the signal preprocessing effect, the SPWVD time-frequency analysis method is used to extract the time-frequency image of the signal, which significantly suppresses the generation of cross terms, improves the signal preprocessing effect of automatic modulation recognition, and improves the signal distinguishability.

[0045] 2. The present invention discloses a signal recognition method based on time-frequency analysis and deep learning under DRFM intermittent sampling. By using position coding and random masking, the time-frequency domain characteristics of the signal obtained under the DRFM intermittent sampling scenario are simulated. The feature extraction capability of the automatic modulation recognition algorithm is improved by feature extraction and time-frequency reconstruction, so that it can be applied to scenarios such as DRFM intermittent sampling.

[0046] 3. This invention discloses a signal recognition method based on time-frequency analysis and deep learning under DRFM intermittent sampling. It employs a pre-training and fine-tuning approach to train an automatic modulation and recognition model for radio fuze signals based on Vision Transformer. The pre-training process uses unprocessed complete time-frequency images, and extracts signal features through time-frequency reconstruction and random masking. Parameter transfer is used to transfer the parameters of the pre-trained model to the fine-tuned model. The fine-tuning process requires only a small number of intermittently sampled time-frequency images for training. This pre-training plus fine-tuning training method improves the training efficiency of the Vision Transformer-based automatic modulation and recognition model for radio fuze signals and significantly reduces training costs.

[0047] 4. The present invention discloses a signal recognition method based on time-frequency analysis and deep learning under DRFM intermittent sampling. It uses SPWVD to extract the time-frequency image of the signal, trains an automatic modulation recognition model of radio fuze signal based on Vision Transformer as a classifier, and combines signal processing algorithm and deep learning algorithm to improve the recognition accuracy and efficiency of radio fuze signal in DRFM intermittent sampling scenario. Attached Figure Description

[0048] Figure 1 This is a flowchart of the signal recognition method based on time-frequency analysis and deep learning under DRFM intermittent sampling of the present invention;

[0049] Figure 2 This is a schematic diagram of the intermittent sampling direct forwarding of the present invention;

[0050] Figure 3 This is a schematic diagram of the intermittent sampling and repeated forwarding of the present invention;

[0051] Figure 4 This is a schematic diagram of the intermittent sampling and cyclic forwarding of the present invention;

[0052] Figure 5 This is the SPWVD time-frequency diagram of the signal preprocessing process in this invention example; wherein, Figure 5 (a) is the SPWVD time-frequency diagram of the complete signal; Figure 5 (b) is the SPWVD time-frequency diagram of the signal obtained in the intermittent sampling direct forwarding mode.

[0053] Figure 6 This is a structural diagram of the automatic modulation and identification model for radio fuze signals based on Vision Transformer of the present invention;

[0054] Figure 7This is the confusion matrix of radio fuze signals with different modulation types identified by the automatic modulation recognition model based on Vision Transformer in this invention example;

[0055] Figure 8 This is a curve showing the change in signal-to-noise ratio of the recognition accuracy of the automatic modulation recognition model for radio fuze signals based on Vision Transformer in this invention example for different modulation types of radio fuze signals. Detailed Implementation

[0056] To better illustrate the purpose and advantages of the present invention, the invention will be further described below in conjunction with the accompanying drawings and examples.

[0057] Example 1:

[0058] To verify the feasibility of this method, the modulation identification of five types of radio fuze signals is used as an example to provide specific steps for this example method. Figure 1 As shown in the figure, this example discloses a signal recognition method based on time-frequency analysis and deep learning under DRFM intermittent sampling. The specific implementation steps are as follows:

[0059] Step 1: The radio fuse jammer based on Digital Radio Frequency Storage (DRFM) acquires the time-frequency truncated radio fuse signal using an intermittent sampling mode according to different jamming and forwarding strategies, and obtains the down-converted target signal through down-conversion processing.

[0060] The radio fuse jammer based on Digital Radio Frequency Storage (DRFM) employs a time-division multiplexing antenna to acquire target signals and relay jamming signals. During the transmission of jamming signals, no signals are received. Therefore, the DRFM jamming equipment operates in an intermittent sampling and jamming relay mode.

[0061] The target signal obtained through intermittent sampling will exhibit time-frequency domain truncation characteristics. The sampling method of intermittent sampling is related to the interference forwarding strategy. There are three DRFM interference forwarding strategies: direct forwarding, repeated forwarding, and cyclic forwarding. Different interference forwarding strategies result in different time-domain signal truncation effects and signal lengths. When the duration of a single sampling is the same as the duration of a single forwarding interference, the signal truncation patterns under different interference forwarding strategies have the following characteristics.

[0062] like Figure 2 As shown, in the intermittent sampling and direct forwarding process, the radio fuse jammer receives the target signal and directly forwards it. For example... Figure 3As shown, during the intermittent sampling and repeated forwarding process, after receiving the target signal, the radio fuse jammer first repeatedly splices the same signal segment multiple times, and then forwards the spliced ​​jamming signal. For example... Figure 4 As shown, during the intermittent sampling and cyclic forwarding process, after receiving the target signal, the radio fuse jammer sequentially splices the previously obtained signals and then forwards the spliced ​​jamming signal. During the cyclic forwarding process, the signal splicing process is repeated periodically, and the signals stored in the previous cycle are not forwarded.

[0063] Depending on the different interference forwarding strategies, the radio fuze receiver acquires the time-frequency truncated target signal through intermittent sampling, and performs down-conversion processing on the target signal to obtain the down-converted target signal.

[0064] Step 2: Using the smoothed pseudo-Wigner-Villi distribution SPWVD, the target signal after frequency conversion processing in Step 1 is windowed and smoothed in both the time and frequency domains to suppress the generation of cross terms, enhance the time-frequency characteristics of the signal, and generate the SPWVD time-frequency image of the signal.

[0065] SPWVD performs time-domain and frequency-domain windowing smoothing on the signal, suppressing the generation of cross terms and enhancing the time-frequency characteristics of the signal.

[0066] For a discrete signal x(n), where n = 0, 1, 2, ..., N-1, its SPWVD operation expression is:

[0067]

[0068] Where t(m) represents the smoothing window function in the time domain and g(k) represents the smoothing window function in the frequency domain, both of which are Hamming windows.

[0069] After obtaining the time series of the radio fuze signal by intermittent sampling, the time-frequency analysis method of Equation (5) is used to perform time-frequency analysis on the time series to obtain the SPWVD time-frequency image of the signal.

[0070] Figure 5 This is the time-frequency diagram of the signal obtained through SPWVD time-frequency analysis, where... Figure 5 (a) is the SPWVD time-frequency diagram of the complete signal. Figure 5 (b) is the SPWVD time-frequency diagram of the signal obtained in the intermittent sampling direct forwarding mode.

[0071] Step 3: Establish an automatic modulation and recognition model for radio fuze signals based on the Vision Transformer backbone network. This model consists of an encoder, a decoder, and an output layer, with the encoder and decoder using the Masked Autoencoders algorithm structure. The time-frequency image is input into the encoder, which performs positional encoding and random masking on the image, extracts features from the time-frequency image using the Transformer structure, and outputs a feature map. The decoder reconstructs the time-frequency image by fitting the difference between the feature map and the original time-frequency image. The feature extraction and image reconstruction capabilities of the Vision Transformer-based automatic modulation and recognition model for radio fuze signals are trained by minimizing the difference between the reconstructed and original time-frequency images. After reconstructing the time-frequency image, the decoder outputs a one-dimensional feature vector to the output layer. The output layer performs dimensionality reduction and classification on the one-dimensional feature vector to obtain a one-dimensional recognition vector representing the modulation type of the radio fuze signal. The training process of the Vision Transformer-based automatic modulation and recognition model for radio fuze signals adopts a pre-training plus fine-tuning training method to reduce the training cost and improve the training efficiency. During pre-training, time-frequency reconstruction training is performed using complete, unlabeled time-frequency images of radio fuze signals, without modulation type identification. The fine-tuning training process first transfers the parameters of the pre-trained Vision Transformer-based automatic modulation identification model of radio fuze signals to the fine-tuning model via parameter transfer, initializing the fine-tuning model. Then, a small number of time-frequency images of radio fuze signals acquired through intermittent sampling are used to train the Vision Transformer-based automatic modulation identification model for modulation type identification.

[0072] Step 3.1: Establish an automatic modulation and identification model for radio fuze signals based on the Vision Transformer backbone network;

[0073] Figure 6 This document describes the network structure and algorithm flow of an automatic modulation and recognition model for radio fuze signals based on Vision Transformer. Based on its network structure, an automatic modulation and recognition model for radio fuze signals based on Vision Transformer is constructed using Python 3.9. The encoder of the automatic modulation and recognition model for radio fuze signals based on Vision Transformer consists of 12 Transformer layers, with 12 multi-head attention heads, while the decoder consists of 8 Transformer layers, with 16 multi-head attention heads.

[0074] Step 3.2: Train the automatic modulation and recognition model of radio fuze signals based on Vision Transformer using pre-training and fine-tuning to reduce the training cost of the model and improve its recognition performance.

[0075] The Vision Transformer-based automatic modulation model for radio fuze signals was trained using a pre-training and fine-tuning approach. The training data used in the pre-training process consisted of unlabeled, continuous, and complete time-frequency images in the time-frequency domain, such as... Figure 5 As shown in (a), since the pre-training process uses unlabeled, complete time-frequency images, modulation type identification of the radio fuze signal is not performed during pre-training. During fine-tuning, the initial parameters of the model are transferred from the pre-trained model. The training data used for fine-tuning is a small amount of labeled time-frequency images obtained through intermittent sampling, such as... Figure 5 As shown in (b).

[0076] Step 4: Input the SPWVD time-frequency diagram from Step 2 into the VisionTransformer-based automatic modulation and identification model for radio fuze signals trained in Step 3. The model sorts the signal modulation type and identifies radio fuze signals of different systems, thereby improving the accuracy and efficiency of radio fuze signal identification in DRFM intermittent sampling scenarios.

[0077] This invention focuses on intermittent sampling direct forwarding interference as its research object. Experiments are conducted to verify the effectiveness and recognition performance of a signal recognition method based on time-frequency analysis and deep learning under intermittent sampling (DRFM) proposed in this invention. Five radio fuze signals are selected for experimental verification. These radio fuze signals include TRIFM (triangular wave linear frequency modulation), SINFM (sine wave frequency modulation), STWFM (sawtooth wave frequency modulation), PSD (pseudo-code phase modulation), and PSPD (pseudo-code phase modulation pulse Doppler). The time-domain waveforms of the radio fuze signals are obtained using intermittent sampling direct forwarding. SPWVD is then used to obtain the time-frequency distribution image of the fuze signals. Experimental data are generated in the MATLAB environment. Specific parameter settings for various radio fuze signals are shown in Table 1. Parameter settings for intermittent sampling direct forwarding are shown in Table 2.

[0078] Table 1. Parameters of radio fuze signals with different modulation types

[0079]

[0080] Table 2 Parameters for intermittent sampling

[0081]

[0082] During pre-training, a complete time-frequency image dataset in the time-frequency domain was used. The parameters of this dataset are shown in Table 1, with a signal-to-noise ratio (SNR) ranging from -15dB to 0dB. Each radio fuze signal has 6000 samples, and no labels were applied to the samples in this dataset. During fine-tuning training, a truncated time-frequency dataset obtained through intermittent sampling was used. The parameters of this dataset are shown in Tables 1 and 2, with a SNR ranging from -15dB to 0dB. Each radio fuze signal has 300 samples, and the training set to test set ratio is 2:1. The learning rate during both pre-training and fine-tuning training was 0.000125, and the number of batches was 32. The number of training epochs during fine-tuning was 50.

[0083] Figure 7 This is the confusion matrix for different modulation types of radio fuze signals identified by the Vision Transformer-based automatic modulation recognition model in this invention example. The time-frequency image data is from the fine-tuning dataset. In the confusion matrix, the values ​​on the diagonal represent the consistency between the modulation type predicted by the model and the actual modulation type of the signal; the larger the diagonal value, the better the model's recognition performance. Figure 7 As shown, the automatic modulation recognition model for radio fuze signals based on Vision Transformer achieves excellent recognition results for radio fuze signals with different modulation types.

[0084] Figure 8 This is a curve showing the accuracy of the automatic modulation recognition model for radio fuze signals based on Vision Transformer in this invention, as a function of signal-to-noise ratio (SNR), for different modulation types of radio fuze signals. The time-frequency image data used is from a validation dataset with an SNR range of -20dB to 0dB and a step size of 2dB. The parameters of this dataset are shown in Tables 1 and 2, and it was not used in the model training process. Figure 8As shown, the automatic modulation recognition model for radio fuze signals based on Vision Transformer consistently maintains a recognition rate of over 90% for triangular wave linear frequency modulated (RFFM) fuze signals. The model's modulation recognition accuracy for pseudo-code phase-modulated (PMM) and pseudo-code PMM pulse Doppler (PDD) fuze signals rapidly improves within the range of -20dB to -16dB, reaching over 95% when the signal-to-noise ratio (SNR) is higher than -16dB. Similarly, the model's recognition accuracy for sinusoidal wave RFFM and sawtooth wave RFFM fuze signals rapidly improves within the range of -20dB to -12dB, reaching over 95% when the SNR is higher than -12dB. By analyzing the recognition accuracy of the automatic modulation recognition model for radio fuze signals based on Vision Transformer under different SNRs for radio fuze signals of different modulation types, the effectiveness and recognition performance of the proposed signal recognition method based on time-frequency analysis and deep learning under DRFM intermittent sampling are verified. Under low signal-to-noise ratio conditions, the signal recognition method based on time-frequency analysis and deep learning proposed in this invention under DRFM intermittent sampling can accurately identify the modulation type of different radio fuze signals.

[0085] The above detailed description further illustrates the purpose, technical solution, and beneficial effects of the invention. It should be understood that the above description is merely a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A signal recognition method based on time-frequency analysis and deep learning under DRFM intermittent sampling, characterized in that: Includes the following steps, Step 1: The radio fuse jammer based on Digital Radio Frequency Storage (DRFM) acquires the time-frequency truncated radio fuse signal using an intermittent sampling mode according to different jamming and forwarding strategies, and obtains the down-converted target signal through down-conversion processing. Step 2: Using smooth pseudo-Wigner-Villi distribution SPWVD, the target signal after frequency conversion processing in Step 1 is windowed and smoothed in both the time and frequency domains to suppress the generation of cross terms, enhance the time-frequency characteristics of the signal, and generate the SPWVD time-frequency image of the signal. Step 3: Establish an automatic modulation and recognition model for radio fuze signals based on the Vision Transformer backbone network. This model consists of an encoder, a decoder, and an output layer, with the encoder and decoder using the Masked Autoencoders algorithm structure. The time-frequency image is input into the encoder, which performs position encoding and random masking on the time-frequency image, extracts features from the time-frequency image through the Transformer structure, and outputs a feature map of the image. The decoder reconstructs the time-frequency image by fitting the difference between the feature map and the original time-frequency image, and trains the feature extraction and image reconstruction capabilities of the Vision Transformer-based automatic modulation and recognition model for radio fuze signals by minimizing the difference between the reconstructed time-frequency image and the original time-frequency image. After reconstructing the time-frequency image, the decoder outputs a one-dimensional feature vector to the output layer. The output layer performs dimensionality reduction and classification on the one-dimensional feature vector to obtain a one-dimensional recognition vector representing the modulation type of the radio fuze signal. The training process of the automatic modulation recognition model for radio fuze signals based on Vision Transformer adopts a pre-training plus fine-tuning training method to reduce the training cost and improve the training efficiency of the model. In the pre-training process, time-frequency reconstruction training is performed using complete, unlabeled time-frequency images of radio fuze signals. Modulation type recognition of radio fuze signals is not performed during this training process. The fine-tuning training process first transfers the parameters of the pre-trained Vision Transformer-based automatic modulation recognition model for radio fuse signals to the fine-tuning model through parameter transfer, thus initializing the fine-tuning model; then, it uses the time-frequency images of the radio fuse signals obtained by intermittent sampling to train the Vision Transformer-based automatic modulation recognition model for modulation type identification. Step 4: Input the SPWVD time-frequency diagram from Step 2 into the Vision Transformer-based automatic modulation and identification model for radio fuze signals trained in Step 3. The model sorts the signal modulation type and identifies radio fuze signals of different systems, thereby improving the accuracy and efficiency of radio fuze signal identification in DRFM intermittent sampling scenarios.

2. The signal recognition method based on time-frequency analysis and deep learning under DRFM intermittent sampling as described in claim 1, characterized in that: The implementation method for step one is as follows: The radio fuse jammer based on Digital Radio Frequency Storage (DRFM) uses a time-division multiplexing antenna to acquire target signals and forward jamming signals; it does not receive signals while transmitting jamming signals. Therefore, during the operation of DRFM jamming equipment, an intermittent sampling and forwarding jamming working mode is formed; The target signal obtained through intermittent sampling exhibits time-frequency domain truncation characteristics. The sampling method of intermittent sampling is related to the jamming forwarding strategy. There are three DRFM jamming forwarding strategies: direct forwarding, repeated forwarding, and cyclic forwarding. Different jamming forwarding strategies result in different time-domain signal truncation effects and signal lengths. The signal truncation patterns under different jamming forwarding strategies have the following characteristics: In the process of intermittent sampling direct forwarding, after receiving the target signal, the radio fuze jammer directly forwards the target signal. At this time, the time-domain signal obtained through intermittent sampling is discontinuous and continuous. In the process of intermittent sampling repeated forwarding, after receiving the target signal, the radio fuze jammer first repeatedly splices the same signal segment multiple times, and then forwards the spliced ​​jamming signal. In the process of intermittent sampling cyclic forwarding, after receiving the target signal, the radio fuze jammer sequentially splices the previously obtained signals, and then forwards the spliced ​​jamming signal. During the cyclic forwarding process, the signal splicing process is repeated periodically, and the signals stored in the previous cycle are not forwarded. Depending on the different interference forwarding strategies, the radio fuze receiver acquires the time-frequency truncated target signal through intermittent sampling, and performs down-conversion processing on the target signal to obtain the down-converted target signal.

3. The signal recognition method based on time-frequency analysis and deep learning under DRFM intermittent sampling as described in claim 2, characterized in that: The second step is implemented as follows: SPWVD performs time-domain and frequency-domain windowing smoothing on the signal, suppressing the generation of cross terms and enhancing the time-frequency characteristics of the signal. For discrete signals ,in Then its SPWVD operation expression is: (1) in The smoothing window function representing the time domain. The smoothing window function representing the frequency domain uses Hamming windows for both the time domain and the frequency domain. After obtaining the time series of the radio fuze signal by intermittent sampling, the time-frequency analysis method of Equation (1) is used to perform time-frequency analysis on the time series to obtain the SPWVD time-frequency image of the signal.

4. The signal recognition method based on time-frequency analysis and deep learning under DRFM intermittent sampling as described in claim 3, characterized in that: The method for implementing step three is as follows: Step 3.1: Establish an automatic modulation and identification model for radio fuze signals based on the Vision Transformer backbone network; The automatic modulation and identification model for radio fuze signals based on Vision Transformer consists of an encoder, a decoder, and an output layer. The encoder and decoder use the Masked Autoencoders algorithm structure and are both composed of cascaded Transformer blocks. The encoder and decoder are connected sequentially. To balance model structural complexity with recognition performance, reduce computational costs, and improve recognition accuracy, the encoder consists of 12 layers of Transformer blocks, with 12 multi-head attention heads; the decoder consists of 8 layers of Transformer blocks, with 16 multi-head attention heads. The Transformer block is entirely based on the attention mechanism, without using any convolutional layers or recurrent neural network layers. A Transformer block contains two normalized Linear Neural Network (LN) layers, one multi-head attention layer, and one multilayer perceptron. The first LN layer is cascaded with the multi-head attention layer to form a residual structure, and the second LN layer is cascaded with the multilayer perceptron to form another residual structure; these two residual structures are connected sequentially. Within the Transformer block, the multi-head attention layer consists of multiple self-attention layers and a fully connected layer. The fully connected layer aggregates the outputs of each self-attention layer to obtain the overall output. Multi-head attention enhances the network's stability and robustness. The multilayer perceptron consists of two fully connected layers, one Gaussian error linear unit (GELU) activation function layer, and two dropout layers. Randomly discarding some features through the dropout layer improves the network's robustness and suppresses overfitting. GELU is a variant of the ReLU linear rectified function. The GELU activation function has better non-linear performance than ReLU, significantly improving model performance. The GELU activation function is expressed as... (2) The output layer is connected to the decoder and consists of a fully connected layer and a Softmax activation function layer. The decoder output passes through the fully connected layer to obtain a one-dimensional probability vector of length N, where N represents the total number of modulation types of the radio fuze signal. The Softmax activation function layer then processes the one-dimensional probability vector to maximize the output probability. The Softmax activation function can transform unnormalized predictions into non-negative numbers that sum to 1. The automatic modulation recognition model for radio fuze signals based on Vision Transformer outputs a one-dimensional recognition vector composed of 0s and 1s representing the modulation type of the radio fuze signal. Step 3.2: Train the automatic modulation and recognition model of radio fuze signals based on Vision Transformer using pre-training and fine-tuning to reduce the training cost of the model and improve its recognition performance. The Vision Transformer-based automatic modulation recognition model for radio fuze signals was trained using a pre-training and fine-tuning approach. The pre-training process used unlabeled, continuous, and complete time-frequency images. The automatic modulation recognition model simulated intermittent sampling by randomly discarding some time-frequency image blocks. During pre-training, the loss function used was the mean squared error (MSE). (3) in Represents the actual sample distribution. The sample distribution representing the prediction; The model's feature extraction capability is trained by fitting the difference between the time-frequency image generated by the decoder and the actual time-frequency image through time-frequency reconstruction. Since the pre-training process uses unlabeled time-frequency images with complete time-frequency domain, modulation type identification of the radio fuze signal is not performed during the pre-training process. During fine-tuning training, the initial parameters of the model are transferred from the pre-trained model; the training data used for fine-tuning the model is a small amount of labeled time-frequency images obtained through intermittent sampling; the loss function used in fine-tuning training is the cross-entropy loss function. (4) in Represents the actual category. Represents the category of the prediction; Cross-entropy loss function is used in multi-class classification tasks to calculate the difference between predicted and true values. The model's ability to identify radio fuze modulation types is trained by fitting the error between the predicted modulation type and the true modulation type of the radio fuze signal from the Vision Transformer-based automatic modulation recognition model. The optimization algorithm used in the pre-training and fine-tuning training processes is AdamW. AdamW improves the generalization performance of the Vision Transformer-based automatic modulation recognition model by increasing regularization weights, while maintaining the model's ability to converge quickly.

5. The signal recognition method based on time-frequency analysis and deep learning under DRFM intermittent sampling as described in claim 4, characterized in that: The specific steps for signal sorting and identification of the SPWVD time-frequency diagram in step two, using the Vision Transformer-based automatic modulation and identification model for radio fuze signals, are as follows: Step 4.1: The encoder of the automatic modulation and recognition model for radio fuze signals based on Vision Transformer segments the time-frequency image of the input radio fuze signal into image blocks of equal size; Step 4.2: The encoder performs position encoding on all image blocks and randomly removes some image blocks; Step 4.3: The encoder extracts features from the remaining image patches using the Transformer structure; Step 4.4: Pad the missing parts of the feature map with zeros according to the position encoding; Step 4.5: The decoder of the automatic modulation and identification model of the radio fuze signal reconstructs the original time-frequency image based on the zero-padded feature map and outputs a one-dimensional feature vector; Step 4.6: The output layer of the automatic modulation and identification model for radio fuze signals performs dimensionality reduction and classification on the one-dimensional feature vector output by the decoder, and maximizes the prediction probability; Step 4.7: The output layer outputs a one-dimensional identification vector representing the modulation type of the radio fuze signal, realizing the sorting and identification of radio fuze signals of different systems.

Images